A versioned research record

What changed

Data additions and corrections from committed repository history.

2026-09-220fcded8 ↗

Update T-Head roadmap and optical capacity evidence

Updated: data/benchmarks.yaml
3 field changes
- assessment: Follow-up through September 21 did not verify a matched independent comparison of current Chinese accelerators against Rubin, GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. A Pengcheng Laboratory preprint adds measured Ascend 910C scientific-computing results against A800/H800 controls, but it is not a model-training or inference comparison and does not use current frontier foreign hardware. Absence of a verified public result is not evidence of incapability. → Follow-up through September 22 did not verify a matched independent comparison of current Chinese accelerators against Rubin, GB300, MI355X, TPU v7, or OpenAI Jalapeño. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. A Pengcheng Laboratory preprint adds measured Ascend 910C scientific-computing results against A800/H800 controls, but it is not a model-training or inference comparison and does not use current frontier foreign hardware. Absence of a verified public result is not evidence of incapability.
- reviewed_at: 2026-09-21 → 2026-09-22
- studies: [{"title":"Ascend 910 scientific-computing study","evidence":"measured","date":"2026-07-22","finding":"Pengcheng Laboratory measured Ascend 910A/B/C across scientific kernels and five applications. In one matched quantum-simulation case, a tuned 910C implementation completed a 30-qubit, 30-layer circuit in 11.4 seconds versus 14.3 seconds on A800 with cuQuantum. The study also reports weaker vector and memory-bound behavior, relies on workload-specific optimization, compares against export-limited A800/H800 rather than current frontier chips, and does not establish AI model-training or inference parity.","source_ids":["ascend-science-2026-07-22"]},{"title":"Biren GLM-5.3-Flash functional validation","evidence":"vendor","date":"2026-09-07","finding":"Vendor reports functional and accuracy checks on Bili 166M with SGLang and BIRENSUPA. Compatibility evidence only; no matched foreign-hardware throughput, latency or power comparison.","source_ids":["biren-glm-2026-09-07"]},{"title":"MLPerf Training / Inference v6.0","evidence":"standardized","date":"2026","finding":"Standardized submitter-run results; no matched latest Chinese-versus-foreign comparison established in this baseline.","source_ids":["mlperf-training","mlperf-inference"]},{"title":"InferenceX","evidence":"measured","date":"Continuous","finding":"Reproducible independent inference runs. Rubin, GB300, MI355X and TPU v7 are represented; equivalent current Chinese results were not verified.","source_ids":["inferencex"]},{"title":"Ascend operational field study","evidence":"measured","date":"2026-07","finding":"A small independent preprint documents deployment friction on older Ascend 910 devices. No matched NVIDIA control.","source_ids":["ascend-field"]},{"title":"HiFloat4 pretraining","evidence":"vendor","date":"2026-04","finding":"Demonstrates training progress on dense and MoE models; does not measure frontier time-to-quality against foreign hardware.","source_ids":["hifloat4"]}] → [{"title":"Ascend 910 scientific-computing study","evidence":"measured","date":"2026-07-22","finding":"Pengcheng Laboratory measured Ascend 910A/B/C across scientific kernels and five applications. In one matched quantum-simulation case, a tuned 910C implementation completed a 30-qubit, 30-layer circuit in 11.4 seconds versus 14.3 seconds on A800 with cuQuantum. The study also reports weaker vector and memory-bound behavior, relies on workload-specific optimization, compares against export-limited A800/H800 rather than current frontier chips, and does not establish AI model-training or inference parity.","source_ids":["ascend-science-2026-07-22"]},{"title":"Biren GLM-5.3-Flash functional validation","evidence":"vendor","date":"2026-09-07","finding":"Vendor reports functional and accuracy checks on Bili 166M with SGLang and BIRENSUPA. Compatibility evidence only; no matched foreign-hardware throughput, latency or power comparison.","source_ids":["biren-glm-2026-09-07"]},{"title":"MLPerf Training / Inference v6.0","evidence":"standardized","date":"2026","finding":"Standardized submitter-run results; no matched latest Chinese-versus-foreign comparison established in this baseline.","source_ids":["mlperf-training","mlperf-inference"]},{"title":"InferenceX","evidence":"measured","date":"Continuous","finding":"Reproducible independent inference runs. The current dashboard names Rubin, GB300, MI355X, TPU v7 and OpenAI Jalapeño systems; equivalent current Chinese results were not verified.","source_ids":["inferencex"]},{"title":"Ascend operational field study","evidence":"measured","date":"2026-07","finding":"A small independent preprint documents deployment friction on older Ascend 910 devices. No matched NVIDIA control.","source_ids":["ascend-field"]},{"title":"HiFloat4 pretraining","evidence":"vendor","date":"2026-04","finding":"Demonstrates training progress on dense and MoE models; does not measure frontier time-to-quality against foreign hardware.","source_ids":["hifloat4"]}]
Added: Ligent Technologies
Updated: Alibaba T-Head
5 field changes
- assessment: T-Head reports broad shipment and customer deployment, but does not disclose the product mix, customer acceptance, utilization, workload performance, power, fabrication or memory provenance. Vendor performance multipliers still require a disclosed baseline and workload. → T-Head reports broad shipment and customer deployment and has accelerated the V900 commercialization target, but does not disclose precision-specific workload performance, power, fabrication, memory provenance, customer acceptance or operating utilization. Vendor performance and cluster-scale claims still require disclosed baselines, topology and reproducible workloads.
- facts: [{"label":"M890 rack system","value":"Alibaba said the Panjiu AL128 packages 128 M890 accelerators in one rack and was available to Chinese enterprise customers through Bailian; operating and customer-acceptance data were not disclosed.","evidence":"reported","as_of":"2026-05-20","source_ids":["m890"]},{"label":"Accelerator roadmap","value":"Alibaba targets Zhenwu V900 for Q3 2027 and J900 for Q3 2028; these are expected milestones, not delivered products, and the reported performance multipliers lack a disclosed workload.","evidence":"reported","as_of":"2026-05-20","source_ids":["m890"]}] → [{"label":"M890 rack system","value":"Alibaba said the Panjiu AL128 packages 128 M890 accelerators in one rack and was available to Chinese enterprise customers through Bailian; operating and customer-acceptance data were not disclosed.","evidence":"reported","as_of":"2026-05-20","source_ids":["m890"]},{"label":"Accelerator roadmap","value":"Alibaba targets Zhenwu V900 for Q3 2027 and J900 for Q3 2028; these are expected milestones, not delivered products, and the reported performance multipliers lack a disclosed workload.","evidence":"reported","as_of":"2026-05-20","source_ids":["m890"]},{"label":"V900 launch update","value":"Alibaba unveiled V900 and now targets mass production and commercial release in Q1 2027, earlier than its May Q3 2027 roadmap; it claims three times M890 performance and clusters of up to 500,000 chips without disclosing workload, precision, topology, power or operating results.","evidence":"reported","as_of":"2026-09-22","source_ids":["alibaba-v900-2026-09-22"]}]
- watch_for: ["Measured cloud utilization","Independent model-serving economics","Memory and manufacturing provenance","Product mix, utilization and acceptance behind reported shipment counts"] → ["Measured cloud utilization","Independent model-serving economics","Memory and manufacturing provenance","Product mix, utilization and acceptance behind reported shipment counts","V900 production, customer acceptance and reproducible cluster-scale workloads"]
- source_ids: ["m890","thead"] → ["m890","thead","alibaba-v900-2026-09-22"]
- reviewed_at: 2026-09-19 → 2026-09-22
Added: Alibaba unveils Zhenwu V900 and advances commercial-release target
Added: Ligent raises HKD 5.67 billion for optical R&D and capacity
Updated: data/research-state.yaml
4 field changes
- last_research_at: 2026-09-21 → 2026-09-22
- last_automated_run_at: 2026-09-21 → 2026-09-22
- coverage: The September 14-20 complete weekly briefing synthesizes six canonical in-period events after source review and bounded English and Chinese gap searches. It separates system roadmaps from delivered operating evidence, commodity DRAM from HBM, module assembly from component sovereignty and specialized scientific workloads from AI model training or inference. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts. → The September 15-22 overlap sweep covered all priority-1 companies and the three oldest-reviewed remaining profiles in English and Chinese, then added today's T-Head roadmap change and newly tracked Ligent financing and capacity evidence. Search-only coverage did not advance other company review dates. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders, undisclosed customer and production metrics, and individual benchmark artifacts.
- runs: [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-16","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","kunlunxin","metax","moore-threads"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 9-16. Added Moore Threads' September 9 primary announcement that JD Cloud plans a 100,000-card full-function GPU cluster. The record treats this as a plan rather than delivery evidence because no schedule, GPU model, completed card count, topology, interconnect, power, recovery or workload result was disclosed. Backfilled Moore Threads' August half-year report: company-reported revenue was RMB 1.736 billion, attributable net loss was RMB 11.56 million, and a RMB 660 million KUAE sales contract was fully delivered and recognized. Its Dense and MoE MFU, scaling-efficiency and effective- training-time figures remain vendor measurements without a matched workload, power or current foreign-hardware comparison. Added Kunlunxin's downloadable FlagOS MiniCPM5-2B deployment image and MetaX's ATRIA Dawn Preview Day-0 adaptation plus reported PyTorch operator coverage as compatibility evidence, not performance or CUDA parity. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the half-year filing is explicitly late-added.","source_ids":["moore-threads-2026-h1-report","moore-threads-jd-cloud-2026-09-09","kunlunxin-minicpm-2026-09-10","metax-atria-2026-09-14","mthreads","kunlunxin","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-17","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","skyverse"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 10-17. Added Reuters' September 17 account of Huawei's new launch targets for the 960DT in Q1 2027 and Ascend 960PR in Q3 2027, plus Huawei's claim that it had shipped more than 1,000 supernodes to more than 370 customers. The record preserves the older dated roadmap and treats the new dates as expected milestones, not delivery evidence. Huawei did not identify customers, state the number of accelerators per shipped system or disclose acceptance, utilization, workload, power, reliability or interconnect-performance results. Added Huawei's September 10 account of China Unicom deploying Ascend compute and a 384-NPU supernode and adapting Ascend 950 for Unicom Cloud; this is named-deployment evidence but remains a vendor account without independent operating metrics. Reviewed Skyverse's current official portfolio and news pages without accepting prototype or production claims from image-only reporting. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["huawei-connect-2026-09-17","huawei-china-unicom-2026-09-10","skyverse","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-18","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","smee"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 11-18. Added Reuters' September 18 report that CXMT plans a NAND R&D production line at its new Beijing plant and has discussed prospective AI-storage customers. The event is recorded as a research and trial- production plan, not commercial output: Reuters could not establish an operating date or a commitment to volume manufacturing, and CXMT did not comment. Re-reviewed Reuters' materially expanded September 17 Huawei report and added a companion event for the company's statements enough AI computing equipment to satisfy domestic demand, Ascend 950 systems have entered commercial use, an Ascend 960 supernode is planned for up to 4,096 processors and Peerium is intended eventually to connect as many as one million. These are reported company claims; Huawei did not disclose capacity, orders, customer acceptance, utilization, workload, power, reliability or interconnect-performance data. Reviewed SMEE's current official product page; no new qualified production milestone was accepted. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-nand-2026-09-18","huawei-connect-2026-09-17","smee","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-19","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","thead"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 12-19. Added an independent teardown of one Eoptolink 400G DR4+2km module that directly identified a Marvell IN010C50 DSP and several Analog Devices power and analog parts. The record excludes the author's inferred laser identification and limits the finding to this anonymous, single sample with no disclosed manufacturing date; it does not establish the bill of materials for current 800G or 1.6T production. The oldest-profile review reopened T-Head's current official page and the existing May Reuters report, then backfilled the materially under-recorded company claims of more than 560,000 Zhenwu units shipped, deployments at more than 400 external customers across 20 industries, the 128-accelerator Panjiu AL128 rack and expected V900 and J900 roadmap dates. Product mix, acceptance, utilization, workload performance, power, fabrication and memory provenance remain undisclosed. RSLaser and SIOM searches again produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["eoptolink-400g-teardown-2026-09-19","m890","thead","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-20","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","zhiqi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 13-20. Added Reuters' September 20 account of CXMT's statement that its fifth-generation DRAM platform entered mass production using quadruple patterning and an 11.95 nm active-area half-pitch, and that two 24 Gb LPDDR5X products made on the platform are in mass production. The record treats these as company claims and preserves Reuters' distinction between gross dies per wafer and the proportion passing final tests: the claimed at-least-50% improvement over the fourth- generation platform does not disclose yield, reliability, customer qualification or output volume. CXMT said Chinese equipment makers contributed to critical process steps but did not name them or establish a fully domestic process flow. The event is a commodity-DRAM milestone, not evidence of HBM volume or performance. RSLaser, SIOM and Zhiqi searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage, which now names Rubin alongside GB300, MI355X and TPU v7. No matched contemporary Chinese accelerator result was verified. Search-only priority coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-g5-dram-2026-09-20","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-21","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","zhiqi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 14-21. Late-added a July 22 Pengcheng Laboratory preprint that measures Ascend 910A/B/C on scientific kernels and five application studies. The record captures a favorable matched quantum-simulation case: a tuned 910C implementation completed a 30-qubit, 30-layer circuit in 11.4 seconds versus 14.3 seconds on an A800 using cuQuantum. It also preserves the paper's less favorable evidence: vector- and memory-bound kernels remained constrained, several comparisons used different algorithms or implementations, and workload-specific optimization was central. The study uses export-limited A800/H800 controls rather than current frontier hardware and does not test AI model training or inference, so it does not change the latest-versus-latest assessment. RSLaser, SIOM and Zhiqi searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against Rubin, GB300, MI355X, TPU v7 or the newly listed Jalapeno systems was verified. Search-only priority coverage did not advance other company review dates. Historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["ascend-science-2026-07-22","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-21","status":"updated","companies":["cxmt","eoptolink","huawei","metax","ymtc"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and created the complete September 14-20 edition without replacing the earlier September 7 period-to-date or complete reports. Audited all six selected in-period events and their original sources, then ran bounded English and Chinese gap searches across accelerators, memory, optical interconnects, software and manufacturing equipment. The synthesis treats Huawei's accelerated 2027 roadmap, reported system scale and Peerium targets as company evidence rather than delivered workload performance, and makes Huawei's inability to meet Chinese demand the clearest capacity constraint. It distinguishes CXMT's fifth-generation commodity-DRAM and LPDDR5X mass-production claims from tested yield and HBM progress, and its planned NAND R&D line from commercial output. It uses the single Eoptolink 400G teardown only as sample-specific component evidence. The pre-period Ascend scientific- computing paper is background because it was public by the period end; it does not enter news_ids or establish model-training or inference parity. Rechecked unchanged MLPerf v6.0 repository heads and current InferenceX coverage. No matched current Chinese result against Rubin, GB300, MI355X or TPU v7 was verified, and no dated forecast changed. Coverage remains incomplete for private and unindexed evidence, tenders, individual benchmark artifacts and undisclosed production or customer metrics.","source_ids":["ascend-science-2026-07-22","cioe-optics-2026-09-11","cxmt-g5-dram-2026-09-20","cxmt-nand-2026-09-18","eoptolink-400g-teardown-2026-09-19","huawei-connect-2026-09-17","inferencex","metax-atria-2026-09-14","mlperf-inference","mlperf-training"]}] → [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-16","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","kunlunxin","metax","moore-threads"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 9-16. Added Moore Threads' September 9 primary announcement that JD Cloud plans a 100,000-card full-function GPU cluster. The record treats this as a plan rather than delivery evidence because no schedule, GPU model, completed card count, topology, interconnect, power, recovery or workload result was disclosed. Backfilled Moore Threads' August half-year report: company-reported revenue was RMB 1.736 billion, attributable net loss was RMB 11.56 million, and a RMB 660 million KUAE sales contract was fully delivered and recognized. Its Dense and MoE MFU, scaling-efficiency and effective- training-time figures remain vendor measurements without a matched workload, power or current foreign-hardware comparison. Added Kunlunxin's downloadable FlagOS MiniCPM5-2B deployment image and MetaX's ATRIA Dawn Preview Day-0 adaptation plus reported PyTorch operator coverage as compatibility evidence, not performance or CUDA parity. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the half-year filing is explicitly late-added.","source_ids":["moore-threads-2026-h1-report","moore-threads-jd-cloud-2026-09-09","kunlunxin-minicpm-2026-09-10","metax-atria-2026-09-14","mthreads","kunlunxin","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-17","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","skyverse"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 10-17. Added Reuters' September 17 account of Huawei's new launch targets for the 960DT in Q1 2027 and Ascend 960PR in Q3 2027, plus Huawei's claim that it had shipped more than 1,000 supernodes to more than 370 customers. The record preserves the older dated roadmap and treats the new dates as expected milestones, not delivery evidence. Huawei did not identify customers, state the number of accelerators per shipped system or disclose acceptance, utilization, workload, power, reliability or interconnect-performance results. Added Huawei's September 10 account of China Unicom deploying Ascend compute and a 384-NPU supernode and adapting Ascend 950 for Unicom Cloud; this is named-deployment evidence but remains a vendor account without independent operating metrics. Reviewed Skyverse's current official portfolio and news pages without accepting prototype or production claims from image-only reporting. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["huawei-connect-2026-09-17","huawei-china-unicom-2026-09-10","skyverse","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-18","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","smee"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 11-18. Added Reuters' September 18 report that CXMT plans a NAND R&D production line at its new Beijing plant and has discussed prospective AI-storage customers. The event is recorded as a research and trial- production plan, not commercial output: Reuters could not establish an operating date or a commitment to volume manufacturing, and CXMT did not comment. Re-reviewed Reuters' materially expanded September 17 Huawei report and added a companion event for the company's statements enough AI computing equipment to satisfy domestic demand, Ascend 950 systems have entered commercial use, an Ascend 960 supernode is planned for up to 4,096 processors and Peerium is intended eventually to connect as many as one million. These are reported company claims; Huawei did not disclose capacity, orders, customer acceptance, utilization, workload, power, reliability or interconnect-performance data. Reviewed SMEE's current official product page; no new qualified production milestone was accepted. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-nand-2026-09-18","huawei-connect-2026-09-17","smee","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-19","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","thead"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 12-19. Added an independent teardown of one Eoptolink 400G DR4+2km module that directly identified a Marvell IN010C50 DSP and several Analog Devices power and analog parts. The record excludes the author's inferred laser identification and limits the finding to this anonymous, single sample with no disclosed manufacturing date; it does not establish the bill of materials for current 800G or 1.6T production. The oldest-profile review reopened T-Head's current official page and the existing May Reuters report, then backfilled the materially under-recorded company claims of more than 560,000 Zhenwu units shipped, deployments at more than 400 external customers across 20 industries, the 128-accelerator Panjiu AL128 rack and expected V900 and J900 roadmap dates. Product mix, acceptance, utilization, workload performance, power, fabrication and memory provenance remain undisclosed. RSLaser and SIOM searches again produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["eoptolink-400g-teardown-2026-09-19","m890","thead","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-20","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","zhiqi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 13-20. Added Reuters' September 20 account of CXMT's statement that its fifth-generation DRAM platform entered mass production using quadruple patterning and an 11.95 nm active-area half-pitch, and that two 24 Gb LPDDR5X products made on the platform are in mass production. The record treats these as company claims and preserves Reuters' distinction between gross dies per wafer and the proportion passing final tests: the claimed at-least-50% improvement over the fourth- generation platform does not disclose yield, reliability, customer qualification or output volume. CXMT said Chinese equipment makers contributed to critical process steps but did not name them or establish a fully domestic process flow. The event is a commodity-DRAM milestone, not evidence of HBM volume or performance. RSLaser, SIOM and Zhiqi searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage, which now names Rubin alongside GB300, MI355X and TPU v7. No matched contemporary Chinese accelerator result was verified. Search-only priority coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-g5-dram-2026-09-20","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-21","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","zhiqi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 14-21. Late-added a July 22 Pengcheng Laboratory preprint that measures Ascend 910A/B/C on scientific kernels and five application studies. The record captures a favorable matched quantum-simulation case: a tuned 910C implementation completed a 30-qubit, 30-layer circuit in 11.4 seconds versus 14.3 seconds on an A800 using cuQuantum. It also preserves the paper's less favorable evidence: vector- and memory-bound kernels remained constrained, several comparisons used different algorithms or implementations, and workload-specific optimization was central. The study uses export-limited A800/H800 controls rather than current frontier hardware and does not test AI model training or inference, so it does not change the latest-versus-latest assessment. RSLaser, SIOM and Zhiqi searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against Rubin, GB300, MI355X, TPU v7 or the newly listed Jalapeno systems was verified. Search-only priority coverage did not advance other company review dates. Historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["ascend-science-2026-07-22","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-21","status":"updated","companies":["cxmt","eoptolink","huawei","metax","ymtc"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and created the complete September 14-20 edition without replacing the earlier September 7 period-to-date or complete reports. Audited all six selected in-period events and their original sources, then ran bounded English and Chinese gap searches across accelerators, memory, optical interconnects, software and manufacturing equipment. The synthesis treats Huawei's accelerated 2027 roadmap, reported system scale and Peerium targets as company evidence rather than delivered workload performance, and makes Huawei's inability to meet Chinese demand the clearest capacity constraint. It distinguishes CXMT's fifth-generation commodity-DRAM and LPDDR5X mass-production claims from tested yield and HBM progress, and its planned NAND R&D line from commercial output. It uses the single Eoptolink 400G teardown only as sample-specific component evidence. The pre-period Ascend scientific- computing paper is background because it was public by the period end; it does not enter news_ids or establish model-training or inference parity. Rechecked unchanged MLPerf v6.0 repository heads and current InferenceX coverage. No matched current Chinese result against Rubin, GB300, MI355X or TPU v7 was verified, and no dated forecast changed. Coverage remains incomplete for private and unindexed evidence, tenders, individual benchmark artifacts and undisclosed production or customer metrics.","source_ids":["ascend-science-2026-07-22","cioe-optics-2026-09-11","cxmt-g5-dram-2026-09-20","cxmt-nand-2026-09-18","eoptolink-400g-teardown-2026-09-19","huawei-connect-2026-09-17","inferencex","metax-atria-2026-09-14","mlperf-inference","mlperf-training"]},{"date":"2026-09-22","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","zhiqi","thead","ligent"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 15-22. Added Reuters' September 22 report that Alibaba unveiled T-Head's Zhenwu V900 and now targets Q1 2027 mass production and commercial release, earlier than the Q3 2027 roadmap reported in May. Preserved that older dated roadmap and recorded the new three-times-M890 performance and 500,000-chip cluster claims as reported targets rather than workload evidence: precision, model, power, topology, recovery, fabrication and memory data were not disclosed. Added Ligent Technologies as a tracked photonics supplier after its HKD 5.67 billion Hong Kong IPO, with planned allocations of about 52.9% of net proceeds to faster optical-product and chip R&D and 25.1% to capacity. The record separates capital raised and a vendor-listed 400G/800G/1.6T portfolio from verified high-speed shipment mix, yield, customer qualification and domestic component provenance. RSLaser, SIOM and Zhiqi searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked the unchanged MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current dashboard; no matched contemporary Chinese result against Rubin, GB300, MI355X, TPU v7 or OpenAI Jalapeño was verified. Search-only priority coverage did not advance other company review dates.","source_ids":["alibaba-v900-2026-09-22","ligent","ligent-ipo-2026-09-22","mlperf-training","mlperf-inference","inferencex"]}]
Added: Alibaba deepens AI push with new chip, bigger model; shares jump 5%
Updated: Continuous inference benchmark dashboard
1 field changes
- reviewed_at: 2026-09-21 → 2026-09-22
Added: China's Ligent jumps 19% in Hong Kong debut as AI demand draws investors
Added: Ligent Technologies high-speed optical product portfolio
Updated: MLPerf Inference v6.0 submissions
1 field changes
- reviewed_at: 2026-09-21 → 2026-09-22
Updated: MLPerf Training v6.0 submissions
1 field changes
- reviewed_at: 2026-09-21 → 2026-09-22

2026-09-210634e6a ↗

Publish September 14-20 weekly briefing

Added: Systems and commodity memory advance; capacity and components still bind
Updated: data/research-state.yaml
2 field changes
- coverage: The September 21 daily review screened all 12 priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. It late-added a measured Pengcheng Laboratory study of Ascend 910 scientific-computing workloads, while separating specialized tuned results against A800/H800 from AI model-training or inference parity. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts; only Huawei, whose underlying evidence was actually reviewed, advanced. → The September 14-20 complete weekly briefing synthesizes six canonical in-period events after source review and bounded English and Chinese gap searches. It separates system roadmaps from delivered operating evidence, commodity DRAM from HBM, module assembly from component sovereignty and specialized scientific workloads from AI model training or inference. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts.
- runs: [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-16","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","kunlunxin","metax","moore-threads"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 9-16. Added Moore Threads' September 9 primary announcement that JD Cloud plans a 100,000-card full-function GPU cluster. The record treats this as a plan rather than delivery evidence because no schedule, GPU model, completed card count, topology, interconnect, power, recovery or workload result was disclosed. Backfilled Moore Threads' August half-year report: company-reported revenue was RMB 1.736 billion, attributable net loss was RMB 11.56 million, and a RMB 660 million KUAE sales contract was fully delivered and recognized. Its Dense and MoE MFU, scaling-efficiency and effective- training-time figures remain vendor measurements without a matched workload, power or current foreign-hardware comparison. Added Kunlunxin's downloadable FlagOS MiniCPM5-2B deployment image and MetaX's ATRIA Dawn Preview Day-0 adaptation plus reported PyTorch operator coverage as compatibility evidence, not performance or CUDA parity. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the half-year filing is explicitly late-added.","source_ids":["moore-threads-2026-h1-report","moore-threads-jd-cloud-2026-09-09","kunlunxin-minicpm-2026-09-10","metax-atria-2026-09-14","mthreads","kunlunxin","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-17","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","skyverse"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 10-17. Added Reuters' September 17 account of Huawei's new launch targets for the 960DT in Q1 2027 and Ascend 960PR in Q3 2027, plus Huawei's claim that it had shipped more than 1,000 supernodes to more than 370 customers. The record preserves the older dated roadmap and treats the new dates as expected milestones, not delivery evidence. Huawei did not identify customers, state the number of accelerators per shipped system or disclose acceptance, utilization, workload, power, reliability or interconnect-performance results. Added Huawei's September 10 account of China Unicom deploying Ascend compute and a 384-NPU supernode and adapting Ascend 950 for Unicom Cloud; this is named-deployment evidence but remains a vendor account without independent operating metrics. Reviewed Skyverse's current official portfolio and news pages without accepting prototype or production claims from image-only reporting. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["huawei-connect-2026-09-17","huawei-china-unicom-2026-09-10","skyverse","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-18","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","smee"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 11-18. Added Reuters' September 18 report that CXMT plans a NAND R&D production line at its new Beijing plant and has discussed prospective AI-storage customers. The event is recorded as a research and trial- production plan, not commercial output: Reuters could not establish an operating date or a commitment to volume manufacturing, and CXMT did not comment. Re-reviewed Reuters' materially expanded September 17 Huawei report and added a companion event for the company's statements enough AI computing equipment to satisfy domestic demand, Ascend 950 systems have entered commercial use, an Ascend 960 supernode is planned for up to 4,096 processors and Peerium is intended eventually to connect as many as one million. These are reported company claims; Huawei did not disclose capacity, orders, customer acceptance, utilization, workload, power, reliability or interconnect-performance data. Reviewed SMEE's current official product page; no new qualified production milestone was accepted. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-nand-2026-09-18","huawei-connect-2026-09-17","smee","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-19","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","thead"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 12-19. Added an independent teardown of one Eoptolink 400G DR4+2km module that directly identified a Marvell IN010C50 DSP and several Analog Devices power and analog parts. The record excludes the author's inferred laser identification and limits the finding to this anonymous, single sample with no disclosed manufacturing date; it does not establish the bill of materials for current 800G or 1.6T production. The oldest-profile review reopened T-Head's current official page and the existing May Reuters report, then backfilled the materially under-recorded company claims of more than 560,000 Zhenwu units shipped, deployments at more than 400 external customers across 20 industries, the 128-accelerator Panjiu AL128 rack and expected V900 and J900 roadmap dates. Product mix, acceptance, utilization, workload performance, power, fabrication and memory provenance remain undisclosed. RSLaser and SIOM searches again produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["eoptolink-400g-teardown-2026-09-19","m890","thead","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-20","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","zhiqi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 13-20. Added Reuters' September 20 account of CXMT's statement that its fifth-generation DRAM platform entered mass production using quadruple patterning and an 11.95 nm active-area half-pitch, and that two 24 Gb LPDDR5X products made on the platform are in mass production. The record treats these as company claims and preserves Reuters' distinction between gross dies per wafer and the proportion passing final tests: the claimed at-least-50% improvement over the fourth- generation platform does not disclose yield, reliability, customer qualification or output volume. CXMT said Chinese equipment makers contributed to critical process steps but did not name them or establish a fully domestic process flow. The event is a commodity-DRAM milestone, not evidence of HBM volume or performance. RSLaser, SIOM and Zhiqi searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage, which now names Rubin alongside GB300, MI355X and TPU v7. No matched contemporary Chinese accelerator result was verified. Search-only priority coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-g5-dram-2026-09-20","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-21","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","zhiqi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 14-21. Late-added a July 22 Pengcheng Laboratory preprint that measures Ascend 910A/B/C on scientific kernels and five application studies. The record captures a favorable matched quantum-simulation case: a tuned 910C implementation completed a 30-qubit, 30-layer circuit in 11.4 seconds versus 14.3 seconds on an A800 using cuQuantum. It also preserves the paper's less favorable evidence: vector- and memory-bound kernels remained constrained, several comparisons used different algorithms or implementations, and workload-specific optimization was central. The study uses export-limited A800/H800 controls rather than current frontier hardware and does not test AI model training or inference, so it does not change the latest-versus-latest assessment. RSLaser, SIOM and Zhiqi searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against Rubin, GB300, MI355X, TPU v7 or the newly listed Jalapeno systems was verified. Search-only priority coverage did not advance other company review dates. Historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["ascend-science-2026-07-22","mlperf-training","mlperf-inference","inferencex"]}] → [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-16","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","kunlunxin","metax","moore-threads"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 9-16. Added Moore Threads' September 9 primary announcement that JD Cloud plans a 100,000-card full-function GPU cluster. The record treats this as a plan rather than delivery evidence because no schedule, GPU model, completed card count, topology, interconnect, power, recovery or workload result was disclosed. Backfilled Moore Threads' August half-year report: company-reported revenue was RMB 1.736 billion, attributable net loss was RMB 11.56 million, and a RMB 660 million KUAE sales contract was fully delivered and recognized. Its Dense and MoE MFU, scaling-efficiency and effective- training-time figures remain vendor measurements without a matched workload, power or current foreign-hardware comparison. Added Kunlunxin's downloadable FlagOS MiniCPM5-2B deployment image and MetaX's ATRIA Dawn Preview Day-0 adaptation plus reported PyTorch operator coverage as compatibility evidence, not performance or CUDA parity. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the half-year filing is explicitly late-added.","source_ids":["moore-threads-2026-h1-report","moore-threads-jd-cloud-2026-09-09","kunlunxin-minicpm-2026-09-10","metax-atria-2026-09-14","mthreads","kunlunxin","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-17","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","skyverse"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 10-17. Added Reuters' September 17 account of Huawei's new launch targets for the 960DT in Q1 2027 and Ascend 960PR in Q3 2027, plus Huawei's claim that it had shipped more than 1,000 supernodes to more than 370 customers. The record preserves the older dated roadmap and treats the new dates as expected milestones, not delivery evidence. Huawei did not identify customers, state the number of accelerators per shipped system or disclose acceptance, utilization, workload, power, reliability or interconnect-performance results. Added Huawei's September 10 account of China Unicom deploying Ascend compute and a 384-NPU supernode and adapting Ascend 950 for Unicom Cloud; this is named-deployment evidence but remains a vendor account without independent operating metrics. Reviewed Skyverse's current official portfolio and news pages without accepting prototype or production claims from image-only reporting. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["huawei-connect-2026-09-17","huawei-china-unicom-2026-09-10","skyverse","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-18","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","smee"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 11-18. Added Reuters' September 18 report that CXMT plans a NAND R&D production line at its new Beijing plant and has discussed prospective AI-storage customers. The event is recorded as a research and trial- production plan, not commercial output: Reuters could not establish an operating date or a commitment to volume manufacturing, and CXMT did not comment. Re-reviewed Reuters' materially expanded September 17 Huawei report and added a companion event for the company's statements enough AI computing equipment to satisfy domestic demand, Ascend 950 systems have entered commercial use, an Ascend 960 supernode is planned for up to 4,096 processors and Peerium is intended eventually to connect as many as one million. These are reported company claims; Huawei did not disclose capacity, orders, customer acceptance, utilization, workload, power, reliability or interconnect-performance data. Reviewed SMEE's current official product page; no new qualified production milestone was accepted. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-nand-2026-09-18","huawei-connect-2026-09-17","smee","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-19","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","thead"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 12-19. Added an independent teardown of one Eoptolink 400G DR4+2km module that directly identified a Marvell IN010C50 DSP and several Analog Devices power and analog parts. The record excludes the author's inferred laser identification and limits the finding to this anonymous, single sample with no disclosed manufacturing date; it does not establish the bill of materials for current 800G or 1.6T production. The oldest-profile review reopened T-Head's current official page and the existing May Reuters report, then backfilled the materially under-recorded company claims of more than 560,000 Zhenwu units shipped, deployments at more than 400 external customers across 20 industries, the 128-accelerator Panjiu AL128 rack and expected V900 and J900 roadmap dates. Product mix, acceptance, utilization, workload performance, power, fabrication and memory provenance remain undisclosed. RSLaser and SIOM searches again produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["eoptolink-400g-teardown-2026-09-19","m890","thead","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-20","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","zhiqi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 13-20. Added Reuters' September 20 account of CXMT's statement that its fifth-generation DRAM platform entered mass production using quadruple patterning and an 11.95 nm active-area half-pitch, and that two 24 Gb LPDDR5X products made on the platform are in mass production. The record treats these as company claims and preserves Reuters' distinction between gross dies per wafer and the proportion passing final tests: the claimed at-least-50% improvement over the fourth- generation platform does not disclose yield, reliability, customer qualification or output volume. CXMT said Chinese equipment makers contributed to critical process steps but did not name them or establish a fully domestic process flow. The event is a commodity-DRAM milestone, not evidence of HBM volume or performance. RSLaser, SIOM and Zhiqi searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage, which now names Rubin alongside GB300, MI355X and TPU v7. No matched contemporary Chinese accelerator result was verified. Search-only priority coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-g5-dram-2026-09-20","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-21","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","zhiqi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 14-21. Late-added a July 22 Pengcheng Laboratory preprint that measures Ascend 910A/B/C on scientific kernels and five application studies. The record captures a favorable matched quantum-simulation case: a tuned 910C implementation completed a 30-qubit, 30-layer circuit in 11.4 seconds versus 14.3 seconds on an A800 using cuQuantum. It also preserves the paper's less favorable evidence: vector- and memory-bound kernels remained constrained, several comparisons used different algorithms or implementations, and workload-specific optimization was central. The study uses export-limited A800/H800 controls rather than current frontier hardware and does not test AI model training or inference, so it does not change the latest-versus-latest assessment. RSLaser, SIOM and Zhiqi searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against Rubin, GB300, MI355X, TPU v7 or the newly listed Jalapeno systems was verified. Search-only priority coverage did not advance other company review dates. Historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["ascend-science-2026-07-22","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-21","status":"updated","companies":["cxmt","eoptolink","huawei","metax","ymtc"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and created the complete September 14-20 edition without replacing the earlier September 7 period-to-date or complete reports. Audited all six selected in-period events and their original sources, then ran bounded English and Chinese gap searches across accelerators, memory, optical interconnects, software and manufacturing equipment. The synthesis treats Huawei's accelerated 2027 roadmap, reported system scale and Peerium targets as company evidence rather than delivered workload performance, and makes Huawei's inability to meet Chinese demand the clearest capacity constraint. It distinguishes CXMT's fifth-generation commodity-DRAM and LPDDR5X mass-production claims from tested yield and HBM progress, and its planned NAND R&D line from commercial output. It uses the single Eoptolink 400G teardown only as sample-specific component evidence. The pre-period Ascend scientific- computing paper is background because it was public by the period end; it does not enter news_ids or establish model-training or inference parity. Rechecked unchanged MLPerf v6.0 repository heads and current InferenceX coverage. No matched current Chinese result against Rubin, GB300, MI355X or TPU v7 was verified, and no dated forecast changed. Coverage remains incomplete for private and unindexed evidence, tenders, individual benchmark artifacts and undisclosed production or customer metrics.","source_ids":["ascend-science-2026-07-22","cioe-optics-2026-09-11","cxmt-g5-dram-2026-09-20","cxmt-nand-2026-09-18","eoptolink-400g-teardown-2026-09-19","huawei-connect-2026-09-17","inferencex","metax-atria-2026-09-14","mlperf-inference","mlperf-training"]}]

2026-09-211af283a ↗

Add Ascend scientific-computing benchmark evidence

Updated: data/benchmarks.yaml
3 field changes
- assessment: Follow-up through September 20 did not verify a matched independent comparison of current Chinese accelerators against Rubin, GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability. → Follow-up through September 21 did not verify a matched independent comparison of current Chinese accelerators against Rubin, GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. A Pengcheng Laboratory preprint adds measured Ascend 910C scientific-computing results against A800/H800 controls, but it is not a model-training or inference comparison and does not use current frontier foreign hardware. Absence of a verified public result is not evidence of incapability.
- reviewed_at: 2026-09-20 → 2026-09-21
- studies: [{"title":"Biren GLM-5.3-Flash functional validation","evidence":"vendor","date":"2026-09-07","finding":"Vendor reports functional and accuracy checks on Bili 166M with SGLang and BIRENSUPA. Compatibility evidence only; no matched foreign-hardware throughput, latency or power comparison.","source_ids":["biren-glm-2026-09-07"]},{"title":"MLPerf Training / Inference v6.0","evidence":"standardized","date":"2026","finding":"Standardized submitter-run results; no matched latest Chinese-versus-foreign comparison established in this baseline.","source_ids":["mlperf-training","mlperf-inference"]},{"title":"InferenceX","evidence":"measured","date":"Continuous","finding":"Reproducible independent inference runs. Rubin, GB300, MI355X and TPU v7 are represented; equivalent current Chinese results were not verified.","source_ids":["inferencex"]},{"title":"Ascend operational field study","evidence":"measured","date":"2026-07","finding":"A small independent preprint documents deployment friction on older Ascend 910 devices. No matched NVIDIA control.","source_ids":["ascend-field"]},{"title":"HiFloat4 pretraining","evidence":"vendor","date":"2026-04","finding":"Demonstrates training progress on dense and MoE models; does not measure frontier time-to-quality against foreign hardware.","source_ids":["hifloat4"]}] → [{"title":"Ascend 910 scientific-computing study","evidence":"measured","date":"2026-07-22","finding":"Pengcheng Laboratory measured Ascend 910A/B/C across scientific kernels and five applications. In one matched quantum-simulation case, a tuned 910C implementation completed a 30-qubit, 30-layer circuit in 11.4 seconds versus 14.3 seconds on A800 with cuQuantum. The study also reports weaker vector and memory-bound behavior, relies on workload-specific optimization, compares against export-limited A800/H800 rather than current frontier chips, and does not establish AI model-training or inference parity.","source_ids":["ascend-science-2026-07-22"]},{"title":"Biren GLM-5.3-Flash functional validation","evidence":"vendor","date":"2026-09-07","finding":"Vendor reports functional and accuracy checks on Bili 166M with SGLang and BIRENSUPA. Compatibility evidence only; no matched foreign-hardware throughput, latency or power comparison.","source_ids":["biren-glm-2026-09-07"]},{"title":"MLPerf Training / Inference v6.0","evidence":"standardized","date":"2026","finding":"Standardized submitter-run results; no matched latest Chinese-versus-foreign comparison established in this baseline.","source_ids":["mlperf-training","mlperf-inference"]},{"title":"InferenceX","evidence":"measured","date":"Continuous","finding":"Reproducible independent inference runs. Rubin, GB300, MI355X and TPU v7 are represented; equivalent current Chinese results were not verified.","source_ids":["inferencex"]},{"title":"Ascend operational field study","evidence":"measured","date":"2026-07","finding":"A small independent preprint documents deployment friction on older Ascend 910 devices. No matched NVIDIA control.","source_ids":["ascend-field"]},{"title":"HiFloat4 pretraining","evidence":"vendor","date":"2026-04","finding":"Demonstrates training progress on dense and MoE models; does not measure frontier time-to-quality against foreign hardware.","source_ids":["hifloat4"]}]
Updated: Huawei
3 field changes
- facts: [{"label":"Domestic supply constraint","value":"Huawei says it cannot produce enough AI computing equipment to meet Chinese demand and is consequently limiting overseas sales; no capacity, order or fulfillment figures were disclosed.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"Ascend 950 commercialization claim","value":"Huawei materials say Ascend 950 systems entered commercial use; configurations, volume, customer acceptance and workload results were not disclosed.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"Scale-up architecture roadmap","value":"Huawei says Peerium is intended to link up to one million processors and an Ascend 960 supernode is planned for up to 4,096 processors; these are architecture targets, not measured results.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"Reported supernode shipments","value":"Huawei says it shipped more than 1,000 supernodes to more than 370 customers; customer, configuration, utilization and workload details were not disclosed.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"2027 accelerator roadmap","value":"Huawei expects the 960DT in Q1 2027 and Ascend 960PR in Q3 2027; these are launch targets, not delivered-product evidence.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"China Unicom deployment","value":"Huawei says China Unicom deployed Ascend compute at scale and a 384-NPU supernode, and is adapting Ascend 950 for Unicom Cloud; independent operating metrics were not disclosed.","evidence":"vendor","as_of":"2026-09-10","source_ids":["huawei-china-unicom-2026-09-10"]},{"label":"PyTorch accelerator integration","value":"Huawei is a PyTorch Foundation Platinum member and co-chairs its Accelerator Integration Working Group; governance participation does not prove workload parity.","evidence":"primary","as_of":"2026-09-09","source_ids":["pytorch-china-2026-09-09"]},{"label":"Atlas 350 / 950PR","value":"112 GB HBM; 1.4 TB/s; 425 TFLOPS BF16; up to 600 W","evidence":"vendor","as_of":"2026-09-11","source_ids":["atlas350"]}] → [{"label":"Scientific-computing measurements","value":"A Pengcheng Laboratory preprint measured Ascend 910A/B/C on five scientific workloads; a tuned 910C quantum-simulation run took 11.4 seconds versus 14.3 seconds on A800, while vector- and memory-bound kernels remained constrained. This is not a model-training or inference result.","evidence":"measured","as_of":"2026-07-22","source_ids":["ascend-science-2026-07-22"]},{"label":"Domestic supply constraint","value":"Huawei says it cannot produce enough AI computing equipment to meet Chinese demand and is consequently limiting overseas sales; no capacity, order or fulfillment figures were disclosed.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"Ascend 950 commercialization claim","value":"Huawei materials say Ascend 950 systems entered commercial use; configurations, volume, customer acceptance and workload results were not disclosed.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"Scale-up architecture roadmap","value":"Huawei says Peerium is intended to link up to one million processors and an Ascend 960 supernode is planned for up to 4,096 processors; these are architecture targets, not measured results.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"Reported supernode shipments","value":"Huawei says it shipped more than 1,000 supernodes to more than 370 customers; customer, configuration, utilization and workload details were not disclosed.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"2027 accelerator roadmap","value":"Huawei expects the 960DT in Q1 2027 and Ascend 960PR in Q3 2027; these are launch targets, not delivered-product evidence.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"China Unicom deployment","value":"Huawei says China Unicom deployed Ascend compute at scale and a 384-NPU supernode, and is adapting Ascend 950 for Unicom Cloud; independent operating metrics were not disclosed.","evidence":"vendor","as_of":"2026-09-10","source_ids":["huawei-china-unicom-2026-09-10"]},{"label":"PyTorch accelerator integration","value":"Huawei is a PyTorch Foundation Platinum member and co-chairs its Accelerator Integration Working Group; governance participation does not prove workload parity.","evidence":"primary","as_of":"2026-09-09","source_ids":["pytorch-china-2026-09-09"]},{"label":"Atlas 350 / 950PR","value":"112 GB HBM; 1.4 TB/s; 425 TFLOPS BF16; up to 600 W","evidence":"vendor","as_of":"2026-09-11","source_ids":["atlas350"]}]
- source_ids: ["ascend","hbm-reuters","pytorch-china-2026-09-09","huawei-connect-2026-09-17","huawei-china-unicom-2026-09-10"] → ["ascend","hbm-reuters","pytorch-china-2026-09-09","huawei-connect-2026-09-17","huawei-china-unicom-2026-09-10","ascend-science-2026-07-22"]
- reviewed_at: 2026-09-18 → 2026-09-21
Added: Pengcheng Laboratory measures Ascend 910 scientific-computing workloads
Updated: data/research-state.yaml
4 field changes
- last_research_at: 2026-09-20 → 2026-09-21
- last_automated_run_at: 2026-09-20 → 2026-09-21
- coverage: The September 20 daily review screened all 12 priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. It added CXMT's reported fifth-generation DRAM and 24 Gb LPDDR5X mass-production milestone, while separating gross dies per wafer from tested yield and commodity DRAM from HBM. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts; only CXMT, whose underlying evidence was actually reviewed, advanced. → The September 21 daily review screened all 12 priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. It late-added a measured Pengcheng Laboratory study of Ascend 910 scientific-computing workloads, while separating specialized tuned results against A800/H800 from AI model-training or inference parity. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts; only Huawei, whose underlying evidence was actually reviewed, advanced.
- runs: [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-16","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","kunlunxin","metax","moore-threads"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 9-16. Added Moore Threads' September 9 primary announcement that JD Cloud plans a 100,000-card full-function GPU cluster. The record treats this as a plan rather than delivery evidence because no schedule, GPU model, completed card count, topology, interconnect, power, recovery or workload result was disclosed. Backfilled Moore Threads' August half-year report: company-reported revenue was RMB 1.736 billion, attributable net loss was RMB 11.56 million, and a RMB 660 million KUAE sales contract was fully delivered and recognized. Its Dense and MoE MFU, scaling-efficiency and effective- training-time figures remain vendor measurements without a matched workload, power or current foreign-hardware comparison. Added Kunlunxin's downloadable FlagOS MiniCPM5-2B deployment image and MetaX's ATRIA Dawn Preview Day-0 adaptation plus reported PyTorch operator coverage as compatibility evidence, not performance or CUDA parity. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the half-year filing is explicitly late-added.","source_ids":["moore-threads-2026-h1-report","moore-threads-jd-cloud-2026-09-09","kunlunxin-minicpm-2026-09-10","metax-atria-2026-09-14","mthreads","kunlunxin","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-17","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","skyverse"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 10-17. Added Reuters' September 17 account of Huawei's new launch targets for the 960DT in Q1 2027 and Ascend 960PR in Q3 2027, plus Huawei's claim that it had shipped more than 1,000 supernodes to more than 370 customers. The record preserves the older dated roadmap and treats the new dates as expected milestones, not delivery evidence. Huawei did not identify customers, state the number of accelerators per shipped system or disclose acceptance, utilization, workload, power, reliability or interconnect-performance results. Added Huawei's September 10 account of China Unicom deploying Ascend compute and a 384-NPU supernode and adapting Ascend 950 for Unicom Cloud; this is named-deployment evidence but remains a vendor account without independent operating metrics. Reviewed Skyverse's current official portfolio and news pages without accepting prototype or production claims from image-only reporting. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["huawei-connect-2026-09-17","huawei-china-unicom-2026-09-10","skyverse","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-18","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","smee"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 11-18. Added Reuters' September 18 report that CXMT plans a NAND R&D production line at its new Beijing plant and has discussed prospective AI-storage customers. The event is recorded as a research and trial- production plan, not commercial output: Reuters could not establish an operating date or a commitment to volume manufacturing, and CXMT did not comment. Re-reviewed Reuters' materially expanded September 17 Huawei report and added a companion event for the company's statements enough AI computing equipment to satisfy domestic demand, Ascend 950 systems have entered commercial use, an Ascend 960 supernode is planned for up to 4,096 processors and Peerium is intended eventually to connect as many as one million. These are reported company claims; Huawei did not disclose capacity, orders, customer acceptance, utilization, workload, power, reliability or interconnect-performance data. Reviewed SMEE's current official product page; no new qualified production milestone was accepted. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-nand-2026-09-18","huawei-connect-2026-09-17","smee","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-19","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","thead"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 12-19. Added an independent teardown of one Eoptolink 400G DR4+2km module that directly identified a Marvell IN010C50 DSP and several Analog Devices power and analog parts. The record excludes the author's inferred laser identification and limits the finding to this anonymous, single sample with no disclosed manufacturing date; it does not establish the bill of materials for current 800G or 1.6T production. The oldest-profile review reopened T-Head's current official page and the existing May Reuters report, then backfilled the materially under-recorded company claims of more than 560,000 Zhenwu units shipped, deployments at more than 400 external customers across 20 industries, the 128-accelerator Panjiu AL128 rack and expected V900 and J900 roadmap dates. Product mix, acceptance, utilization, workload performance, power, fabrication and memory provenance remain undisclosed. RSLaser and SIOM searches again produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["eoptolink-400g-teardown-2026-09-19","m890","thead","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-20","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","zhiqi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 13-20. Added Reuters' September 20 account of CXMT's statement that its fifth-generation DRAM platform entered mass production using quadruple patterning and an 11.95 nm active-area half-pitch, and that two 24 Gb LPDDR5X products made on the platform are in mass production. The record treats these as company claims and preserves Reuters' distinction between gross dies per wafer and the proportion passing final tests: the claimed at-least-50% improvement over the fourth- generation platform does not disclose yield, reliability, customer qualification or output volume. CXMT said Chinese equipment makers contributed to critical process steps but did not name them or establish a fully domestic process flow. The event is a commodity-DRAM milestone, not evidence of HBM volume or performance. RSLaser, SIOM and Zhiqi searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage, which now names Rubin alongside GB300, MI355X and TPU v7. No matched contemporary Chinese accelerator result was verified. Search-only priority coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-g5-dram-2026-09-20","mlperf-training","mlperf-inference","inferencex"]}] → [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-16","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","kunlunxin","metax","moore-threads"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 9-16. Added Moore Threads' September 9 primary announcement that JD Cloud plans a 100,000-card full-function GPU cluster. The record treats this as a plan rather than delivery evidence because no schedule, GPU model, completed card count, topology, interconnect, power, recovery or workload result was disclosed. Backfilled Moore Threads' August half-year report: company-reported revenue was RMB 1.736 billion, attributable net loss was RMB 11.56 million, and a RMB 660 million KUAE sales contract was fully delivered and recognized. Its Dense and MoE MFU, scaling-efficiency and effective- training-time figures remain vendor measurements without a matched workload, power or current foreign-hardware comparison. Added Kunlunxin's downloadable FlagOS MiniCPM5-2B deployment image and MetaX's ATRIA Dawn Preview Day-0 adaptation plus reported PyTorch operator coverage as compatibility evidence, not performance or CUDA parity. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the half-year filing is explicitly late-added.","source_ids":["moore-threads-2026-h1-report","moore-threads-jd-cloud-2026-09-09","kunlunxin-minicpm-2026-09-10","metax-atria-2026-09-14","mthreads","kunlunxin","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-17","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","skyverse"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 10-17. Added Reuters' September 17 account of Huawei's new launch targets for the 960DT in Q1 2027 and Ascend 960PR in Q3 2027, plus Huawei's claim that it had shipped more than 1,000 supernodes to more than 370 customers. The record preserves the older dated roadmap and treats the new dates as expected milestones, not delivery evidence. Huawei did not identify customers, state the number of accelerators per shipped system or disclose acceptance, utilization, workload, power, reliability or interconnect-performance results. Added Huawei's September 10 account of China Unicom deploying Ascend compute and a 384-NPU supernode and adapting Ascend 950 for Unicom Cloud; this is named-deployment evidence but remains a vendor account without independent operating metrics. Reviewed Skyverse's current official portfolio and news pages without accepting prototype or production claims from image-only reporting. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["huawei-connect-2026-09-17","huawei-china-unicom-2026-09-10","skyverse","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-18","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","smee"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 11-18. Added Reuters' September 18 report that CXMT plans a NAND R&D production line at its new Beijing plant and has discussed prospective AI-storage customers. The event is recorded as a research and trial- production plan, not commercial output: Reuters could not establish an operating date or a commitment to volume manufacturing, and CXMT did not comment. Re-reviewed Reuters' materially expanded September 17 Huawei report and added a companion event for the company's statements enough AI computing equipment to satisfy domestic demand, Ascend 950 systems have entered commercial use, an Ascend 960 supernode is planned for up to 4,096 processors and Peerium is intended eventually to connect as many as one million. These are reported company claims; Huawei did not disclose capacity, orders, customer acceptance, utilization, workload, power, reliability or interconnect-performance data. Reviewed SMEE's current official product page; no new qualified production milestone was accepted. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-nand-2026-09-18","huawei-connect-2026-09-17","smee","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-19","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","thead"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 12-19. Added an independent teardown of one Eoptolink 400G DR4+2km module that directly identified a Marvell IN010C50 DSP and several Analog Devices power and analog parts. The record excludes the author's inferred laser identification and limits the finding to this anonymous, single sample with no disclosed manufacturing date; it does not establish the bill of materials for current 800G or 1.6T production. The oldest-profile review reopened T-Head's current official page and the existing May Reuters report, then backfilled the materially under-recorded company claims of more than 560,000 Zhenwu units shipped, deployments at more than 400 external customers across 20 industries, the 128-accelerator Panjiu AL128 rack and expected V900 and J900 roadmap dates. Product mix, acceptance, utilization, workload performance, power, fabrication and memory provenance remain undisclosed. RSLaser and SIOM searches again produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["eoptolink-400g-teardown-2026-09-19","m890","thead","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-20","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","zhiqi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 13-20. Added Reuters' September 20 account of CXMT's statement that its fifth-generation DRAM platform entered mass production using quadruple patterning and an 11.95 nm active-area half-pitch, and that two 24 Gb LPDDR5X products made on the platform are in mass production. The record treats these as company claims and preserves Reuters' distinction between gross dies per wafer and the proportion passing final tests: the claimed at-least-50% improvement over the fourth- generation platform does not disclose yield, reliability, customer qualification or output volume. CXMT said Chinese equipment makers contributed to critical process steps but did not name them or establish a fully domestic process flow. The event is a commodity-DRAM milestone, not evidence of HBM volume or performance. RSLaser, SIOM and Zhiqi searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage, which now names Rubin alongside GB300, MI355X and TPU v7. No matched contemporary Chinese accelerator result was verified. Search-only priority coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-g5-dram-2026-09-20","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-21","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","zhiqi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 14-21. Late-added a July 22 Pengcheng Laboratory preprint that measures Ascend 910A/B/C on scientific kernels and five application studies. The record captures a favorable matched quantum-simulation case: a tuned 910C implementation completed a 30-qubit, 30-layer circuit in 11.4 seconds versus 14.3 seconds on an A800 using cuQuantum. It also preserves the paper's less favorable evidence: vector- and memory-bound kernels remained constrained, several comparisons used different algorithms or implementations, and workload-specific optimization was central. The study uses export-limited A800/H800 controls rather than current frontier hardware and does not test AI model training or inference, so it does not change the latest-versus-latest assessment. RSLaser, SIOM and Zhiqi searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against Rubin, GB300, MI355X, TPU v7 or the newly listed Jalapeno systems was verified. Search-only priority coverage did not advance other company review dates. Historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["ascend-science-2026-07-22","mlperf-training","mlperf-inference","inferencex"]}]
Added: Ascend to Science: Exploration of AI Chips for Scientific Computing
Updated: Continuous inference benchmark dashboard
1 field changes
- reviewed_at: 2026-09-20 → 2026-09-21
Updated: MLPerf Inference v6.0 submissions
1 field changes
- reviewed_at: 2026-09-20 → 2026-09-21
Updated: MLPerf Training v6.0 submissions
1 field changes
- reviewed_at: 2026-09-20 → 2026-09-21

2026-09-2039ea331 ↗

Update CXMT fifth-generation DRAM evidence

Updated: data/benchmarks.yaml
3 field changes
- assessment: Follow-up through September 19 did not verify a matched independent comparison of current Chinese accelerators against GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability. → Follow-up through September 20 did not verify a matched independent comparison of current Chinese accelerators against Rubin, GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability.
- reviewed_at: 2026-09-19 → 2026-09-20
- studies: [{"title":"Biren GLM-5.3-Flash functional validation","evidence":"vendor","date":"2026-09-07","finding":"Vendor reports functional and accuracy checks on Bili 166M with SGLang and BIRENSUPA. Compatibility evidence only; no matched foreign-hardware throughput, latency or power comparison.","source_ids":["biren-glm-2026-09-07"]},{"title":"MLPerf Training / Inference v6.0","evidence":"standardized","date":"2026","finding":"Standardized submitter-run results; no matched latest Chinese-versus-foreign comparison established in this baseline.","source_ids":["mlperf-training","mlperf-inference"]},{"title":"InferenceX","evidence":"measured","date":"Continuous","finding":"Reproducible independent inference runs. GB300, MI355X and TPU v7 are represented; equivalent current Chinese results were not verified.","source_ids":["inferencex"]},{"title":"Ascend operational field study","evidence":"measured","date":"2026-07","finding":"A small independent preprint documents deployment friction on older Ascend 910 devices. No matched NVIDIA control.","source_ids":["ascend-field"]},{"title":"HiFloat4 pretraining","evidence":"vendor","date":"2026-04","finding":"Demonstrates training progress on dense and MoE models; does not measure frontier time-to-quality against foreign hardware.","source_ids":["hifloat4"]}] → [{"title":"Biren GLM-5.3-Flash functional validation","evidence":"vendor","date":"2026-09-07","finding":"Vendor reports functional and accuracy checks on Bili 166M with SGLang and BIRENSUPA. Compatibility evidence only; no matched foreign-hardware throughput, latency or power comparison.","source_ids":["biren-glm-2026-09-07"]},{"title":"MLPerf Training / Inference v6.0","evidence":"standardized","date":"2026","finding":"Standardized submitter-run results; no matched latest Chinese-versus-foreign comparison established in this baseline.","source_ids":["mlperf-training","mlperf-inference"]},{"title":"InferenceX","evidence":"measured","date":"Continuous","finding":"Reproducible independent inference runs. Rubin, GB300, MI355X and TPU v7 are represented; equivalent current Chinese results were not verified.","source_ids":["inferencex"]},{"title":"Ascend operational field study","evidence":"measured","date":"2026-07","finding":"A small independent preprint documents deployment friction on older Ascend 910 devices. No matched NVIDIA control.","source_ids":["ascend-field"]},{"title":"HiFloat4 pretraining","evidence":"vendor","date":"2026-04","finding":"Demonstrates training progress on dense and MoE models; does not measure frontier time-to-quality against foreign hardware.","source_ids":["hifloat4"]}]
Updated: CXMT
5 field changes
- assessment: Adequate domestic HBM volume and frontier HBM performance are separate milestones. The reported NAND project is an R&D-line plan, not evidence of qualified commercial output. → Fifth-generation commodity DRAM is company-reported in mass production, but adequate domestic HBM volume and frontier HBM performance remain separate milestones. The reported NAND project is an R&D-line plan, not evidence of qualified commercial output.
- facts: [{"label":"NAND expansion plan","value":"Reuters reports CXMT plans a Beijing NAND R&D production line and has discussed prospective AI-storage customers; no operating date or commercial-production commitment was established.","evidence":"reported","as_of":"2026-09-18","source_ids":["cxmt-nand-2026-09-18"]}] → [{"label":"Fifth-generation DRAM platform","value":"CXMT says its fifth-generation platform entered mass production using quadruple patterning and an 11.95 nm active-area half-pitch; independent yield and reliability data were not disclosed.","evidence":"reported","as_of":"2026-09-20","source_ids":["cxmt-g5-dram-2026-09-20"]},{"label":"LPDDR5X production","value":"Two 24 Gb LPDDR5X products are company-reported in mass production, with at least 50% more gross dies per wafer than the fourth-generation platform on an 8 Gb baseline; gross dies are potential dies before defective units are excluded, not tested yield.","evidence":"reported","as_of":"2026-09-20","source_ids":["cxmt-g5-dram-2026-09-20"]},{"label":"NAND expansion plan","value":"Reuters reports CXMT plans a Beijing NAND R&D production line and has discussed prospective AI-storage customers; no operating date or commercial-production commitment was established.","evidence":"reported","as_of":"2026-09-18","source_ids":["cxmt-nand-2026-09-18"]}]
- watch_for: ["Customer-qualified HBM shipments","Stack yield and volume","Accelerators shipping with verified domestic DRAM","NAND trial-line operation and transition to qualified volume production"] → ["Customer-qualified HBM shipments","Stack yield and volume","Accelerators shipping with verified domestic DRAM","Independent G5 yield, reliability and customer qualification","Named domestic-equipment contributions to the G5 process flow","NAND trial-line operation and transition to qualified volume production"]
- source_ids: ["cxmt-hbm","cxmt","cxmt-nand-2026-09-18"] → ["cxmt-hbm","cxmt","cxmt-nand-2026-09-18","cxmt-g5-dram-2026-09-20"]
- reviewed_at: 2026-09-18 → 2026-09-20
Added: CXMT says its fifth-generation DRAM platform and 24 Gb LPDDR5X entered mass production
Updated: data/research-state.yaml
4 field changes
- last_research_at: 2026-09-19 → 2026-09-20
- last_automated_run_at: 2026-09-19 → 2026-09-20
- coverage: The September 19 daily review screened all 12 priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. It added a single-sample independent teardown of an Eoptolink 400G module and backfilled T-Head's reported Zhenwu shipment, customer, rack-system and roadmap details from an existing Reuters source. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts; only Eoptolink and T-Head, whose underlying evidence was actually reviewed, advanced. → The September 20 daily review screened all 12 priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. It added CXMT's reported fifth-generation DRAM and 24 Gb LPDDR5X mass-production milestone, while separating gross dies per wafer from tested yield and commodity DRAM from HBM. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts; only CXMT, whose underlying evidence was actually reviewed, advanced.
- runs: [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-16","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","kunlunxin","metax","moore-threads"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 9-16. Added Moore Threads' September 9 primary announcement that JD Cloud plans a 100,000-card full-function GPU cluster. The record treats this as a plan rather than delivery evidence because no schedule, GPU model, completed card count, topology, interconnect, power, recovery or workload result was disclosed. Backfilled Moore Threads' August half-year report: company-reported revenue was RMB 1.736 billion, attributable net loss was RMB 11.56 million, and a RMB 660 million KUAE sales contract was fully delivered and recognized. Its Dense and MoE MFU, scaling-efficiency and effective- training-time figures remain vendor measurements without a matched workload, power or current foreign-hardware comparison. Added Kunlunxin's downloadable FlagOS MiniCPM5-2B deployment image and MetaX's ATRIA Dawn Preview Day-0 adaptation plus reported PyTorch operator coverage as compatibility evidence, not performance or CUDA parity. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the half-year filing is explicitly late-added.","source_ids":["moore-threads-2026-h1-report","moore-threads-jd-cloud-2026-09-09","kunlunxin-minicpm-2026-09-10","metax-atria-2026-09-14","mthreads","kunlunxin","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-17","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","skyverse"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 10-17. Added Reuters' September 17 account of Huawei's new launch targets for the 960DT in Q1 2027 and Ascend 960PR in Q3 2027, plus Huawei's claim that it had shipped more than 1,000 supernodes to more than 370 customers. The record preserves the older dated roadmap and treats the new dates as expected milestones, not delivery evidence. Huawei did not identify customers, state the number of accelerators per shipped system or disclose acceptance, utilization, workload, power, reliability or interconnect-performance results. Added Huawei's September 10 account of China Unicom deploying Ascend compute and a 384-NPU supernode and adapting Ascend 950 for Unicom Cloud; this is named-deployment evidence but remains a vendor account without independent operating metrics. Reviewed Skyverse's current official portfolio and news pages without accepting prototype or production claims from image-only reporting. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["huawei-connect-2026-09-17","huawei-china-unicom-2026-09-10","skyverse","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-18","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","smee"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 11-18. Added Reuters' September 18 report that CXMT plans a NAND R&D production line at its new Beijing plant and has discussed prospective AI-storage customers. The event is recorded as a research and trial- production plan, not commercial output: Reuters could not establish an operating date or a commitment to volume manufacturing, and CXMT did not comment. Re-reviewed Reuters' materially expanded September 17 Huawei report and added a companion event for the company's statements enough AI computing equipment to satisfy domestic demand, Ascend 950 systems have entered commercial use, an Ascend 960 supernode is planned for up to 4,096 processors and Peerium is intended eventually to connect as many as one million. These are reported company claims; Huawei did not disclose capacity, orders, customer acceptance, utilization, workload, power, reliability or interconnect-performance data. Reviewed SMEE's current official product page; no new qualified production milestone was accepted. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-nand-2026-09-18","huawei-connect-2026-09-17","smee","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-19","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","thead"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 12-19. Added an independent teardown of one Eoptolink 400G DR4+2km module that directly identified a Marvell IN010C50 DSP and several Analog Devices power and analog parts. The record excludes the author's inferred laser identification and limits the finding to this anonymous, single sample with no disclosed manufacturing date; it does not establish the bill of materials for current 800G or 1.6T production. The oldest-profile review reopened T-Head's current official page and the existing May Reuters report, then backfilled the materially under-recorded company claims of more than 560,000 Zhenwu units shipped, deployments at more than 400 external customers across 20 industries, the 128-accelerator Panjiu AL128 rack and expected V900 and J900 roadmap dates. Product mix, acceptance, utilization, workload performance, power, fabrication and memory provenance remain undisclosed. RSLaser and SIOM searches again produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["eoptolink-400g-teardown-2026-09-19","m890","thead","mlperf-training","mlperf-inference","inferencex"]}] → [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-16","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","kunlunxin","metax","moore-threads"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 9-16. Added Moore Threads' September 9 primary announcement that JD Cloud plans a 100,000-card full-function GPU cluster. The record treats this as a plan rather than delivery evidence because no schedule, GPU model, completed card count, topology, interconnect, power, recovery or workload result was disclosed. Backfilled Moore Threads' August half-year report: company-reported revenue was RMB 1.736 billion, attributable net loss was RMB 11.56 million, and a RMB 660 million KUAE sales contract was fully delivered and recognized. Its Dense and MoE MFU, scaling-efficiency and effective- training-time figures remain vendor measurements without a matched workload, power or current foreign-hardware comparison. Added Kunlunxin's downloadable FlagOS MiniCPM5-2B deployment image and MetaX's ATRIA Dawn Preview Day-0 adaptation plus reported PyTorch operator coverage as compatibility evidence, not performance or CUDA parity. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the half-year filing is explicitly late-added.","source_ids":["moore-threads-2026-h1-report","moore-threads-jd-cloud-2026-09-09","kunlunxin-minicpm-2026-09-10","metax-atria-2026-09-14","mthreads","kunlunxin","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-17","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","skyverse"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 10-17. Added Reuters' September 17 account of Huawei's new launch targets for the 960DT in Q1 2027 and Ascend 960PR in Q3 2027, plus Huawei's claim that it had shipped more than 1,000 supernodes to more than 370 customers. The record preserves the older dated roadmap and treats the new dates as expected milestones, not delivery evidence. Huawei did not identify customers, state the number of accelerators per shipped system or disclose acceptance, utilization, workload, power, reliability or interconnect-performance results. Added Huawei's September 10 account of China Unicom deploying Ascend compute and a 384-NPU supernode and adapting Ascend 950 for Unicom Cloud; this is named-deployment evidence but remains a vendor account without independent operating metrics. Reviewed Skyverse's current official portfolio and news pages without accepting prototype or production claims from image-only reporting. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["huawei-connect-2026-09-17","huawei-china-unicom-2026-09-10","skyverse","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-18","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","smee"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 11-18. Added Reuters' September 18 report that CXMT plans a NAND R&D production line at its new Beijing plant and has discussed prospective AI-storage customers. The event is recorded as a research and trial- production plan, not commercial output: Reuters could not establish an operating date or a commitment to volume manufacturing, and CXMT did not comment. Re-reviewed Reuters' materially expanded September 17 Huawei report and added a companion event for the company's statements enough AI computing equipment to satisfy domestic demand, Ascend 950 systems have entered commercial use, an Ascend 960 supernode is planned for up to 4,096 processors and Peerium is intended eventually to connect as many as one million. These are reported company claims; Huawei did not disclose capacity, orders, customer acceptance, utilization, workload, power, reliability or interconnect-performance data. Reviewed SMEE's current official product page; no new qualified production milestone was accepted. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-nand-2026-09-18","huawei-connect-2026-09-17","smee","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-19","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","thead"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 12-19. Added an independent teardown of one Eoptolink 400G DR4+2km module that directly identified a Marvell IN010C50 DSP and several Analog Devices power and analog parts. The record excludes the author's inferred laser identification and limits the finding to this anonymous, single sample with no disclosed manufacturing date; it does not establish the bill of materials for current 800G or 1.6T production. The oldest-profile review reopened T-Head's current official page and the existing May Reuters report, then backfilled the materially under-recorded company claims of more than 560,000 Zhenwu units shipped, deployments at more than 400 external customers across 20 industries, the 128-accelerator Panjiu AL128 rack and expected V900 and J900 roadmap dates. Product mix, acceptance, utilization, workload performance, power, fabrication and memory provenance remain undisclosed. RSLaser and SIOM searches again produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["eoptolink-400g-teardown-2026-09-19","m890","thead","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-20","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","zhiqi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 13-20. Added Reuters' September 20 account of CXMT's statement that its fifth-generation DRAM platform entered mass production using quadruple patterning and an 11.95 nm active-area half-pitch, and that two 24 Gb LPDDR5X products made on the platform are in mass production. The record treats these as company claims and preserves Reuters' distinction between gross dies per wafer and the proportion passing final tests: the claimed at-least-50% improvement over the fourth- generation platform does not disclose yield, reliability, customer qualification or output volume. CXMT said Chinese equipment makers contributed to critical process steps but did not name them or establish a fully domestic process flow. The event is a commodity-DRAM milestone, not evidence of HBM volume or performance. RSLaser, SIOM and Zhiqi searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage, which now names Rubin alongside GB300, MI355X and TPU v7. No matched contemporary Chinese accelerator result was verified. Search-only priority coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-g5-dram-2026-09-20","mlperf-training","mlperf-inference","inferencex"]}]
Added: China's CXMT says new memory-chip platform enters mass production
Updated: Continuous inference benchmark dashboard
1 field changes
- reviewed_at: 2026-09-19 → 2026-09-20
Updated: MLPerf Inference v6.0 submissions
1 field changes
- reviewed_at: 2026-09-19 → 2026-09-20
Updated: MLPerf Training v6.0 submissions
1 field changes
- reviewed_at: 2026-09-19 → 2026-09-20

2026-09-197efd00b ↗

Update optics dependency and T-Head scale evidence

Updated: data/benchmarks.yaml
2 field changes
- assessment: Follow-up through September 18 did not verify a matched independent comparison of current Chinese accelerators against GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability. → Follow-up through September 19 did not verify a matched independent comparison of current Chinese accelerators against GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability.
- reviewed_at: 2026-09-18 → 2026-09-19
Updated: Eoptolink
4 field changes
- assessment: A clear example of global Chinese product competitiveness without total component sovereignty. → A clear example of global Chinese product competitiveness without total component sovereignty. A September 2026 independent teardown of one 400G module identified US DSP and analog/power components, but the anonymous single-sample inspection does not establish the bill of materials for current 800G or 1.6T production.
- facts: [{"label":"1.6T shipment progression","value":"Company reports higher second-quarter than first-quarter 2026 shipments; faster second-half growth is an expectation, not a realized result.","evidence":"reported","as_of":"2026-09-11","source_ids":["cioe-optics-2026-09-11"]}] → [{"label":"1.6T shipment progression","value":"Company reports higher second-quarter than first-quarter 2026 shipments; faster second-half growth is an expectation, not a realized result.","evidence":"reported","as_of":"2026-09-11","source_ids":["cioe-optics-2026-09-11"]},{"label":"400G module component inspection","value":"An independent teardown of one Eoptolink 400G DR4+2km module identified a Marvell IN010C50 DSP and multiple Analog Devices power and analog parts; the module's manufacturing date and representativeness were not established.","evidence":"measured","as_of":"2026-09-19","source_ids":["eoptolink-400g-teardown-2026-09-19"]}]
- source_ids: ["eop-dsp","eoptolink","cioe-optics-2026-09-11"] → ["eop-dsp","eoptolink","cioe-optics-2026-09-11","eoptolink-400g-teardown-2026-09-19"]
- reviewed_at: 2026-09-12 → 2026-09-19
Updated: Alibaba T-Head
6 field changes
- assessment: Vendor performance multipliers require a disclosed baseline and workload. → T-Head reports broad shipment and customer deployment, but does not disclose the product mix, customer acceptance, utilization, workload performance, power, fabrication or memory provenance. Vendor performance multipliers still require a disclosed baseline and workload.
- key_fact.label: Product announcement → Reported deployment scale
- key_fact.value: Zhenwu M890 announced May 2026 → More than 560,000 Zhenwu units shipped; over 400 external customers across 20 industries
- facts: [] → [{"label":"M890 rack system","value":"Alibaba said the Panjiu AL128 packages 128 M890 accelerators in one rack and was available to Chinese enterprise customers through Bailian; operating and customer-acceptance data were not disclosed.","evidence":"reported","as_of":"2026-05-20","source_ids":["m890"]},{"label":"Accelerator roadmap","value":"Alibaba targets Zhenwu V900 for Q3 2027 and J900 for Q3 2028; these are expected milestones, not delivered products, and the reported performance multipliers lack a disclosed workload.","evidence":"reported","as_of":"2026-05-20","source_ids":["m890"]}]
- watch_for: ["Measured cloud utilization","Independent model-serving economics","Memory and manufacturing provenance"] → ["Measured cloud utilization","Independent model-serving economics","Memory and manufacturing provenance","Product mix, utilization and acceptance behind reported shipment counts"]
- reviewed_at: 2026-09-11 → 2026-09-19
Added: Single-module teardown identifies US components in an Eoptolink 400G transceiver
Updated: Alibaba announces Zhenwu M890
2 field changes
- summary: A new cloud-oriented accelerator enters Alibaba’s product roadmap. → Alibaba announced the Zhenwu M890 and Panjiu AL128 rack system. T-Head said it had shipped more than 560,000 Zhenwu units and that more than 400 external customers across 20 industries had deployed the chips; the disclosure did not break out product generations or operating metrics.
- impact: Watch measured deployment economics rather than undisclosed vendor multipliers. → The shipment and customer counts are evidence of reported deployment scale, not independent workload performance or sovereign supply. Watch measured utilization and economics rather than undisclosed vendor multipliers.
Updated: data/research-state.yaml
4 field changes
- last_research_at: 2026-09-18 → 2026-09-19
- last_automated_run_at: 2026-09-18 → 2026-09-19
- coverage: The September 18 daily review screened all 12 priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. It added CXMT's reported NAND R&D-line plan and added a companion Huawei event covering its September 17 domestic supply shortfall, Ascend 950 commercialization claim and scale-up roadmap. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts; only CXMT, Huawei and SMEE, whose reporting or primary pages were actually reviewed, advanced. → The September 19 daily review screened all 12 priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. It added a single-sample independent teardown of an Eoptolink 400G module and backfilled T-Head's reported Zhenwu shipment, customer, rack-system and roadmap details from an existing Reuters source. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts; only Eoptolink and T-Head, whose underlying evidence was actually reviewed, advanced.
- runs: [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-16","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","kunlunxin","metax","moore-threads"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 9-16. Added Moore Threads' September 9 primary announcement that JD Cloud plans a 100,000-card full-function GPU cluster. The record treats this as a plan rather than delivery evidence because no schedule, GPU model, completed card count, topology, interconnect, power, recovery or workload result was disclosed. Backfilled Moore Threads' August half-year report: company-reported revenue was RMB 1.736 billion, attributable net loss was RMB 11.56 million, and a RMB 660 million KUAE sales contract was fully delivered and recognized. Its Dense and MoE MFU, scaling-efficiency and effective- training-time figures remain vendor measurements without a matched workload, power or current foreign-hardware comparison. Added Kunlunxin's downloadable FlagOS MiniCPM5-2B deployment image and MetaX's ATRIA Dawn Preview Day-0 adaptation plus reported PyTorch operator coverage as compatibility evidence, not performance or CUDA parity. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the half-year filing is explicitly late-added.","source_ids":["moore-threads-2026-h1-report","moore-threads-jd-cloud-2026-09-09","kunlunxin-minicpm-2026-09-10","metax-atria-2026-09-14","mthreads","kunlunxin","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-17","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","skyverse"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 10-17. Added Reuters' September 17 account of Huawei's new launch targets for the 960DT in Q1 2027 and Ascend 960PR in Q3 2027, plus Huawei's claim that it had shipped more than 1,000 supernodes to more than 370 customers. The record preserves the older dated roadmap and treats the new dates as expected milestones, not delivery evidence. Huawei did not identify customers, state the number of accelerators per shipped system or disclose acceptance, utilization, workload, power, reliability or interconnect-performance results. Added Huawei's September 10 account of China Unicom deploying Ascend compute and a 384-NPU supernode and adapting Ascend 950 for Unicom Cloud; this is named-deployment evidence but remains a vendor account without independent operating metrics. Reviewed Skyverse's current official portfolio and news pages without accepting prototype or production claims from image-only reporting. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["huawei-connect-2026-09-17","huawei-china-unicom-2026-09-10","skyverse","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-18","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","smee"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 11-18. Added Reuters' September 18 report that CXMT plans a NAND R&D production line at its new Beijing plant and has discussed prospective AI-storage customers. The event is recorded as a research and trial- production plan, not commercial output: Reuters could not establish an operating date or a commitment to volume manufacturing, and CXMT did not comment. Re-reviewed Reuters' materially expanded September 17 Huawei report and added a companion event for the company's statements enough AI computing equipment to satisfy domestic demand, Ascend 950 systems have entered commercial use, an Ascend 960 supernode is planned for up to 4,096 processors and Peerium is intended eventually to connect as many as one million. These are reported company claims; Huawei did not disclose capacity, orders, customer acceptance, utilization, workload, power, reliability or interconnect-performance data. Reviewed SMEE's current official product page; no new qualified production milestone was accepted. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-nand-2026-09-18","huawei-connect-2026-09-17","smee","mlperf-training","mlperf-inference","inferencex"]}] → [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-16","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","kunlunxin","metax","moore-threads"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 9-16. Added Moore Threads' September 9 primary announcement that JD Cloud plans a 100,000-card full-function GPU cluster. The record treats this as a plan rather than delivery evidence because no schedule, GPU model, completed card count, topology, interconnect, power, recovery or workload result was disclosed. Backfilled Moore Threads' August half-year report: company-reported revenue was RMB 1.736 billion, attributable net loss was RMB 11.56 million, and a RMB 660 million KUAE sales contract was fully delivered and recognized. Its Dense and MoE MFU, scaling-efficiency and effective- training-time figures remain vendor measurements without a matched workload, power or current foreign-hardware comparison. Added Kunlunxin's downloadable FlagOS MiniCPM5-2B deployment image and MetaX's ATRIA Dawn Preview Day-0 adaptation plus reported PyTorch operator coverage as compatibility evidence, not performance or CUDA parity. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the half-year filing is explicitly late-added.","source_ids":["moore-threads-2026-h1-report","moore-threads-jd-cloud-2026-09-09","kunlunxin-minicpm-2026-09-10","metax-atria-2026-09-14","mthreads","kunlunxin","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-17","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","skyverse"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 10-17. Added Reuters' September 17 account of Huawei's new launch targets for the 960DT in Q1 2027 and Ascend 960PR in Q3 2027, plus Huawei's claim that it had shipped more than 1,000 supernodes to more than 370 customers. The record preserves the older dated roadmap and treats the new dates as expected milestones, not delivery evidence. Huawei did not identify customers, state the number of accelerators per shipped system or disclose acceptance, utilization, workload, power, reliability or interconnect-performance results. Added Huawei's September 10 account of China Unicom deploying Ascend compute and a 384-NPU supernode and adapting Ascend 950 for Unicom Cloud; this is named-deployment evidence but remains a vendor account without independent operating metrics. Reviewed Skyverse's current official portfolio and news pages without accepting prototype or production claims from image-only reporting. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["huawei-connect-2026-09-17","huawei-china-unicom-2026-09-10","skyverse","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-18","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","smee"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 11-18. Added Reuters' September 18 report that CXMT plans a NAND R&D production line at its new Beijing plant and has discussed prospective AI-storage customers. The event is recorded as a research and trial- production plan, not commercial output: Reuters could not establish an operating date or a commitment to volume manufacturing, and CXMT did not comment. Re-reviewed Reuters' materially expanded September 17 Huawei report and added a companion event for the company's statements enough AI computing equipment to satisfy domestic demand, Ascend 950 systems have entered commercial use, an Ascend 960 supernode is planned for up to 4,096 processors and Peerium is intended eventually to connect as many as one million. These are reported company claims; Huawei did not disclose capacity, orders, customer acceptance, utilization, workload, power, reliability or interconnect-performance data. Reviewed SMEE's current official product page; no new qualified production milestone was accepted. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-nand-2026-09-18","huawei-connect-2026-09-17","smee","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-19","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","thead"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 12-19. Added an independent teardown of one Eoptolink 400G DR4+2km module that directly identified a Marvell IN010C50 DSP and several Analog Devices power and analog parts. The record excludes the author's inferred laser identification and limits the finding to this anonymous, single sample with no disclosed manufacturing date; it does not establish the bill of materials for current 800G or 1.6T production. The oldest-profile review reopened T-Head's current official page and the existing May Reuters report, then backfilled the materially under-recorded company claims of more than 560,000 Zhenwu units shipped, deployments at more than 400 external customers across 20 industries, the 128-accelerator Panjiu AL128 rack and expected V900 and J900 roadmap dates. Product mix, acceptance, utilization, workload performance, power, fabrication and memory provenance remain undisclosed. RSLaser and SIOM searches again produced no retrievable new primary evidence, so their review dates did not advance. Rechecked unchanged MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["eoptolink-400g-teardown-2026-09-19","m890","thead","mlperf-training","mlperf-inference","inferencex"]}]
Added: Teardown of an Eoptolink 400G DR4+2km optical transceiver
Updated: Continuous inference benchmark dashboard
1 field changes
- reviewed_at: 2026-09-18 → 2026-09-19
Updated: Zhenwu M890 announcement
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-19
Updated: MLPerf Inference v6.0 submissions
1 field changes
- reviewed_at: 2026-09-18 → 2026-09-19
Updated: MLPerf Training v6.0 submissions
1 field changes
- reviewed_at: 2026-09-18 → 2026-09-19
Updated: T-Head chip portfolio
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-19

2026-09-18d1a1784 ↗

Update memory and Ascend supply evidence

Updated: data/benchmarks.yaml
2 field changes
- assessment: Follow-up through September 17 did not verify a matched independent comparison of current Chinese accelerators against GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability. → Follow-up through September 18 did not verify a matched independent comparison of current Chinese accelerators against GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability.
- reviewed_at: 2026-09-17 → 2026-09-18
Updated: CXMT
5 field changes
- assessment: Adequate domestic HBM volume and frontier HBM performance are separate milestones. → Adequate domestic HBM volume and frontier HBM performance are separate milestones. The reported NAND project is an R&D-line plan, not evidence of qualified commercial output.
- facts: [] → [{"label":"NAND expansion plan","value":"Reuters reports CXMT plans a Beijing NAND R&D production line and has discussed prospective AI-storage customers; no operating date or commercial-production commitment was established.","evidence":"reported","as_of":"2026-09-18","source_ids":["cxmt-nand-2026-09-18"]}]
- watch_for: ["Customer-qualified HBM shipments","Stack yield and volume","Accelerators shipping with verified domestic DRAM"] → ["Customer-qualified HBM shipments","Stack yield and volume","Accelerators shipping with verified domestic DRAM","NAND trial-line operation and transition to qualified volume production"]
- source_ids: ["cxmt-hbm","cxmt"] → ["cxmt-hbm","cxmt","cxmt-nand-2026-09-18"]
- reviewed_at: 2026-09-11 → 2026-09-18
Updated: Huawei
4 field changes
- assessment: Huawei reports meaningful system shipments and a named carrier deployment, but neither disclosure supplies independently verified utilization or workload performance. Chinese design and deployment also do not establish independence in HBM or fabrication equipment. → Huawei reports meaningful system shipments, commercial Ascend 950 use and a named carrier deployment, but also says production cannot satisfy domestic demand. The disclosures do not supply independently verified utilization or workload performance, and Chinese design and deployment do not establish independence in HBM or fabrication equipment.
- facts: [{"label":"Reported supernode shipments","value":"Huawei says it shipped more than 1,000 supernodes to more than 370 customers; customer, configuration, utilization and workload details were not disclosed.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"2027 accelerator roadmap","value":"Huawei expects the 960DT in Q1 2027 and Ascend 960PR in Q3 2027; these are launch targets, not delivered-product evidence.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"China Unicom deployment","value":"Huawei says China Unicom deployed Ascend compute at scale and a 384-NPU supernode, and is adapting Ascend 950 for Unicom Cloud; independent operating metrics were not disclosed.","evidence":"vendor","as_of":"2026-09-10","source_ids":["huawei-china-unicom-2026-09-10"]},{"label":"PyTorch accelerator integration","value":"Huawei is a PyTorch Foundation Platinum member and co-chairs its Accelerator Integration Working Group; governance participation does not prove workload parity.","evidence":"primary","as_of":"2026-09-09","source_ids":["pytorch-china-2026-09-09"]},{"label":"Atlas 350 / 950PR","value":"112 GB HBM; 1.4 TB/s; 425 TFLOPS BF16; up to 600 W","evidence":"vendor","as_of":"2026-09-11","source_ids":["atlas350"]}] → [{"label":"Domestic supply constraint","value":"Huawei says it cannot produce enough AI computing equipment to meet Chinese demand and is consequently limiting overseas sales; no capacity, order or fulfillment figures were disclosed.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"Ascend 950 commercialization claim","value":"Huawei materials say Ascend 950 systems entered commercial use; configurations, volume, customer acceptance and workload results were not disclosed.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"Scale-up architecture roadmap","value":"Huawei says Peerium is intended to link up to one million processors and an Ascend 960 supernode is planned for up to 4,096 processors; these are architecture targets, not measured results.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"Reported supernode shipments","value":"Huawei says it shipped more than 1,000 supernodes to more than 370 customers; customer, configuration, utilization and workload details were not disclosed.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"2027 accelerator roadmap","value":"Huawei expects the 960DT in Q1 2027 and Ascend 960PR in Q3 2027; these are launch targets, not delivered-product evidence.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"China Unicom deployment","value":"Huawei says China Unicom deployed Ascend compute at scale and a 384-NPU supernode, and is adapting Ascend 950 for Unicom Cloud; independent operating metrics were not disclosed.","evidence":"vendor","as_of":"2026-09-10","source_ids":["huawei-china-unicom-2026-09-10"]},{"label":"PyTorch accelerator integration","value":"Huawei is a PyTorch Foundation Platinum member and co-chairs its Accelerator Integration Working Group; governance participation does not prove workload parity.","evidence":"primary","as_of":"2026-09-09","source_ids":["pytorch-china-2026-09-09"]},{"label":"Atlas 350 / 950PR","value":"112 GB HBM; 1.4 TB/s; 425 TFLOPS BF16; up to 600 W","evidence":"vendor","as_of":"2026-09-11","source_ids":["atlas350"]}]
- watch_for: ["Sustained time-to-quality on large training runs","Customer utilization and recovery statistics","950DT deliveries and verified memory provenance","960DT and 960PR delivery against the new 2027 launch targets"] → ["Sustained time-to-quality on large training runs","Customer utilization and recovery statistics","950DT deliveries and verified memory provenance","960DT and 960PR delivery against the new 2027 launch targets","Capacity, order fulfillment and verified performance of commercial Ascend 950 systems"]
- reviewed_at: 2026-09-17 → 2026-09-18
Updated: SMEE
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-18
Added: CXMT reportedly prepares to enter NAND with a Beijing R&D production line
Added: Huawei reports domestic supply constraints and commercial Ascend 950 systems
Updated: data/research-state.yaml
4 field changes
- last_research_at: 2026-09-17 → 2026-09-18
- last_automated_run_at: 2026-09-17 → 2026-09-18
- coverage: The September 17 daily review screened all 12 priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. It added Huawei's new 2027 accelerator roadmap, company-reported supernode shipment scale and a named China Unicom deployment. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts; only Huawei and Skyverse, whose current reporting or primary pages were actually reviewed, advanced. → The September 18 daily review screened all 12 priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. It added CXMT's reported NAND R&D-line plan and added a companion Huawei event covering its September 17 domestic supply shortfall, Ascend 950 commercialization claim and scale-up roadmap. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts; only CXMT, Huawei and SMEE, whose reporting or primary pages were actually reviewed, advanced.
- runs: [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-16","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","kunlunxin","metax","moore-threads"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 9-16. Added Moore Threads' September 9 primary announcement that JD Cloud plans a 100,000-card full-function GPU cluster. The record treats this as a plan rather than delivery evidence because no schedule, GPU model, completed card count, topology, interconnect, power, recovery or workload result was disclosed. Backfilled Moore Threads' August half-year report: company-reported revenue was RMB 1.736 billion, attributable net loss was RMB 11.56 million, and a RMB 660 million KUAE sales contract was fully delivered and recognized. Its Dense and MoE MFU, scaling-efficiency and effective- training-time figures remain vendor measurements without a matched workload, power or current foreign-hardware comparison. Added Kunlunxin's downloadable FlagOS MiniCPM5-2B deployment image and MetaX's ATRIA Dawn Preview Day-0 adaptation plus reported PyTorch operator coverage as compatibility evidence, not performance or CUDA parity. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the half-year filing is explicitly late-added.","source_ids":["moore-threads-2026-h1-report","moore-threads-jd-cloud-2026-09-09","kunlunxin-minicpm-2026-09-10","metax-atria-2026-09-14","mthreads","kunlunxin","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-17","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","skyverse"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 10-17. Added Reuters' September 17 account of Huawei's new launch targets for the 960DT in Q1 2027 and Ascend 960PR in Q3 2027, plus Huawei's claim that it had shipped more than 1,000 supernodes to more than 370 customers. The record preserves the older dated roadmap and treats the new dates as expected milestones, not delivery evidence. Huawei did not identify customers, state the number of accelerators per shipped system or disclose acceptance, utilization, workload, power, reliability or interconnect-performance results. Added Huawei's September 10 account of China Unicom deploying Ascend compute and a 384-NPU supernode and adapting Ascend 950 for Unicom Cloud; this is named-deployment evidence but remains a vendor account without independent operating metrics. Reviewed Skyverse's current official portfolio and news pages without accepting prototype or production claims from image-only reporting. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["huawei-connect-2026-09-17","huawei-china-unicom-2026-09-10","skyverse","mlperf-training","mlperf-inference","inferencex"]}] → [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-16","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","kunlunxin","metax","moore-threads"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 9-16. Added Moore Threads' September 9 primary announcement that JD Cloud plans a 100,000-card full-function GPU cluster. The record treats this as a plan rather than delivery evidence because no schedule, GPU model, completed card count, topology, interconnect, power, recovery or workload result was disclosed. Backfilled Moore Threads' August half-year report: company-reported revenue was RMB 1.736 billion, attributable net loss was RMB 11.56 million, and a RMB 660 million KUAE sales contract was fully delivered and recognized. Its Dense and MoE MFU, scaling-efficiency and effective- training-time figures remain vendor measurements without a matched workload, power or current foreign-hardware comparison. Added Kunlunxin's downloadable FlagOS MiniCPM5-2B deployment image and MetaX's ATRIA Dawn Preview Day-0 adaptation plus reported PyTorch operator coverage as compatibility evidence, not performance or CUDA parity. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the half-year filing is explicitly late-added.","source_ids":["moore-threads-2026-h1-report","moore-threads-jd-cloud-2026-09-09","kunlunxin-minicpm-2026-09-10","metax-atria-2026-09-14","mthreads","kunlunxin","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-17","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","skyverse"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 10-17. Added Reuters' September 17 account of Huawei's new launch targets for the 960DT in Q1 2027 and Ascend 960PR in Q3 2027, plus Huawei's claim that it had shipped more than 1,000 supernodes to more than 370 customers. The record preserves the older dated roadmap and treats the new dates as expected milestones, not delivery evidence. Huawei did not identify customers, state the number of accelerators per shipped system or disclose acceptance, utilization, workload, power, reliability or interconnect-performance results. Added Huawei's September 10 account of China Unicom deploying Ascend compute and a 384-NPU supernode and adapting Ascend 950 for Unicom Cloud; this is named-deployment evidence but remains a vendor account without independent operating metrics. Reviewed Skyverse's current official portfolio and news pages without accepting prototype or production claims from image-only reporting. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["huawei-connect-2026-09-17","huawei-china-unicom-2026-09-10","skyverse","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-18","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","smee"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 11-18. Added Reuters' September 18 report that CXMT plans a NAND R&D production line at its new Beijing plant and has discussed prospective AI-storage customers. The event is recorded as a research and trial- production plan, not commercial output: Reuters could not establish an operating date or a commitment to volume manufacturing, and CXMT did not comment. Re-reviewed Reuters' materially expanded September 17 Huawei report and added a companion event for the company's statements enough AI computing equipment to satisfy domestic demand, Ascend 950 systems have entered commercial use, an Ascend 960 supernode is planned for up to 4,096 processors and Peerium is intended eventually to connect as many as one million. These are reported company claims; Huawei did not disclose capacity, orders, customer acceptance, utilization, workload, power, reliability or interconnect-performance data. Reviewed SMEE's current official product page; no new qualified production milestone was accepted. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only coverage did not advance other company review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["cxmt-nand-2026-09-18","huawei-connect-2026-09-17","smee","mlperf-training","mlperf-inference","inferencex"]}]
Added: CXMT prepares NAND research line as memory shortage persists
Updated: Huawei says AI chip demand outstrips supply and details its accelerator roadmap
2 field changes
- title: Huawei sets 2027 launch dates for two AI chips and reports supernode shipments → Huawei says AI chip demand outstrips supply and details its accelerator roadmap
- reviewed_at: 2026-09-17 → 2026-09-18
Updated: Continuous inference benchmark dashboard
1 field changes
- reviewed_at: 2026-09-17 → 2026-09-18
Updated: MLPerf Inference v6.0 submissions
1 field changes
- reviewed_at: 2026-09-17 → 2026-09-18
Updated: MLPerf Training v6.0 submissions
1 field changes
- reviewed_at: 2026-09-17 → 2026-09-18
Updated: Lithography equipment portfolio
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-18

2026-09-17da4c7ee ↗

Update Huawei roadmap and deployment evidence

Updated: data/benchmarks.yaml
2 field changes
- assessment: Follow-up through September 16 did not verify a matched independent comparison of current Chinese accelerators against GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability. → Follow-up through September 17 did not verify a matched independent comparison of current Chinese accelerators against GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability.
- reviewed_at: 2026-09-16 → 2026-09-17
Updated: Huawei
5 field changes
- assessment: Chinese design and deployment do not establish independence in HBM or fabrication equipment. → Huawei reports meaningful system shipments and a named carrier deployment, but neither disclosure supplies independently verified utilization or workload performance. Chinese design and deployment also do not establish independence in HBM or fabrication equipment.
- facts: [{"label":"PyTorch accelerator integration","value":"Huawei is a PyTorch Foundation Platinum member and co-chairs its Accelerator Integration Working Group; governance participation does not prove workload parity.","evidence":"primary","as_of":"2026-09-09","source_ids":["pytorch-china-2026-09-09"]},{"label":"Atlas 350 / 950PR","value":"112 GB HBM; 1.4 TB/s; 425 TFLOPS BF16; up to 600 W","evidence":"vendor","as_of":"2026-09-11","source_ids":["atlas350"]}] → [{"label":"Reported supernode shipments","value":"Huawei says it shipped more than 1,000 supernodes to more than 370 customers; customer, configuration, utilization and workload details were not disclosed.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"2027 accelerator roadmap","value":"Huawei expects the 960DT in Q1 2027 and Ascend 960PR in Q3 2027; these are launch targets, not delivered-product evidence.","evidence":"reported","as_of":"2026-09-17","source_ids":["huawei-connect-2026-09-17"]},{"label":"China Unicom deployment","value":"Huawei says China Unicom deployed Ascend compute at scale and a 384-NPU supernode, and is adapting Ascend 950 for Unicom Cloud; independent operating metrics were not disclosed.","evidence":"vendor","as_of":"2026-09-10","source_ids":["huawei-china-unicom-2026-09-10"]},{"label":"PyTorch accelerator integration","value":"Huawei is a PyTorch Foundation Platinum member and co-chairs its Accelerator Integration Working Group; governance participation does not prove workload parity.","evidence":"primary","as_of":"2026-09-09","source_ids":["pytorch-china-2026-09-09"]},{"label":"Atlas 350 / 950PR","value":"112 GB HBM; 1.4 TB/s; 425 TFLOPS BF16; up to 600 W","evidence":"vendor","as_of":"2026-09-11","source_ids":["atlas350"]}]
- watch_for: ["Sustained time-to-quality on large training runs","Customer utilization and recovery statistics","950DT deliveries and verified memory provenance"] → ["Sustained time-to-quality on large training runs","Customer utilization and recovery statistics","950DT deliveries and verified memory provenance","960DT and 960PR delivery against the new 2027 launch targets"]
- source_ids: ["ascend","hbm-reuters","pytorch-china-2026-09-09"] → ["ascend","hbm-reuters","pytorch-china-2026-09-09","huawei-connect-2026-09-17","huawei-china-unicom-2026-09-10"]
- reviewed_at: 2026-09-13 → 2026-09-17
Updated: Skyverse
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-17
Added: Huawei reports China Unicom deployment and Ascend 950 adaptation work
Added: Huawei updates its 2027 accelerator roadmap and reports supernode shipments
Updated: data/research-state.yaml
4 field changes
- last_research_at: 2026-09-16 → 2026-09-17
- last_automated_run_at: 2026-09-16 → 2026-09-17
- coverage: The September 16 daily review screened all 12 priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Current primary disclosures produced software-compatibility updates for Kunlunxin and MetaX, a planned 100,000-GPU JD Cloud cluster for Moore Threads, and a late-added Moore Threads half-year filing with commercial-scale evidence. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts; only the three profiles whose current primary pages or filing were actually reviewed advanced. → The September 17 daily review screened all 12 priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. It added Huawei's new 2027 accelerator roadmap, company-reported supernode shipment scale and a named China Unicom deployment. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts; only Huawei and Skyverse, whose current reporting or primary pages were actually reviewed, advanced.
- runs: [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-16","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","kunlunxin","metax","moore-threads"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 9-16. Added Moore Threads' September 9 primary announcement that JD Cloud plans a 100,000-card full-function GPU cluster. The record treats this as a plan rather than delivery evidence because no schedule, GPU model, completed card count, topology, interconnect, power, recovery or workload result was disclosed. Backfilled Moore Threads' August half-year report: company-reported revenue was RMB 1.736 billion, attributable net loss was RMB 11.56 million, and a RMB 660 million KUAE sales contract was fully delivered and recognized. Its Dense and MoE MFU, scaling-efficiency and effective- training-time figures remain vendor measurements without a matched workload, power or current foreign-hardware comparison. Added Kunlunxin's downloadable FlagOS MiniCPM5-2B deployment image and MetaX's ATRIA Dawn Preview Day-0 adaptation plus reported PyTorch operator coverage as compatibility evidence, not performance or CUDA parity. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the half-year filing is explicitly late-added.","source_ids":["moore-threads-2026-h1-report","moore-threads-jd-cloud-2026-09-09","kunlunxin-minicpm-2026-09-10","metax-atria-2026-09-14","mthreads","kunlunxin","mlperf-training","mlperf-inference","inferencex"]}] → [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-16","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","kunlunxin","metax","moore-threads"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 9-16. Added Moore Threads' September 9 primary announcement that JD Cloud plans a 100,000-card full-function GPU cluster. The record treats this as a plan rather than delivery evidence because no schedule, GPU model, completed card count, topology, interconnect, power, recovery or workload result was disclosed. Backfilled Moore Threads' August half-year report: company-reported revenue was RMB 1.736 billion, attributable net loss was RMB 11.56 million, and a RMB 660 million KUAE sales contract was fully delivered and recognized. Its Dense and MoE MFU, scaling-efficiency and effective- training-time figures remain vendor measurements without a matched workload, power or current foreign-hardware comparison. Added Kunlunxin's downloadable FlagOS MiniCPM5-2B deployment image and MetaX's ATRIA Dawn Preview Day-0 adaptation plus reported PyTorch operator coverage as compatibility evidence, not performance or CUDA parity. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the half-year filing is explicitly late-added.","source_ids":["moore-threads-2026-h1-report","moore-threads-jd-cloud-2026-09-09","kunlunxin-minicpm-2026-09-10","metax-atria-2026-09-14","mthreads","kunlunxin","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-17","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","rslaser","siom","skyverse"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 10-17. Added Reuters' September 17 account of Huawei's new launch targets for the 960DT in Q1 2027 and Ascend 960PR in Q3 2027, plus Huawei's claim that it had shipped more than 1,000 supernodes to more than 370 customers. The record preserves the older dated roadmap and treats the new dates as expected milestones, not delivery evidence. Huawei did not identify customers, state the number of accelerators per shipped system or disclose acceptance, utilization, workload, power, reliability or interconnect-performance results. Added Huawei's September 10 account of China Unicom deploying Ascend compute and a 384-NPU supernode and adapting Ascend 950 for Unicom Cloud; this is named-deployment evidence but remains a vendor account without independent operating metrics. Reviewed Skyverse's current official portfolio and news pages without accepting prototype or production claims from image-only reporting. RSLaser and SIOM searches produced no retrievable new primary evidence, so their review dates did not advance. Rechecked MLPerf Training and Inference v6.0 repository heads, the closed inference submitter directory and InferenceX's current public coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates; historical reports, corrections, forecasts and evidence labels were preserved.","source_ids":["huawei-connect-2026-09-17","huawei-china-unicom-2026-09-10","skyverse","mlperf-training","mlperf-inference","inferencex"]}]
Added: Huawei and China Unicom deepen cooperation around Ascend infrastructure
Added: Huawei sets 2027 launch dates for two AI chips and reports supernode shipments
Updated: Continuous inference benchmark dashboard
1 field changes
- reviewed_at: 2026-09-16 → 2026-09-17
Updated: MLPerf Inference v6.0 submissions
1 field changes
- reviewed_at: 2026-09-16 → 2026-09-17
Updated: MLPerf Training v6.0 submissions
1 field changes
- reviewed_at: 2026-09-16 → 2026-09-17
Updated: Inspection and metrology portfolio
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-17

2026-09-16bdfe641 ↗

Update accelerator scale and software evidence

Updated: data/benchmarks.yaml
2 field changes
- assessment: Follow-up through September 15 did not verify a matched independent comparison of current Chinese accelerators against GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability. → Follow-up through September 16 did not verify a matched independent comparison of current Chinese accelerators against GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability.
- reviewed_at: 2026-09-15 → 2026-09-16
Updated: Kunlunxin
4 field changes
- assessment: Vertical optimization may deliver useful wins before broad ecosystem parity. → Kunlunxin is showing faster model-support cadence through a downloadable FlagOS image for MiniCPM5-2B. The disclosure establishes deployment compatibility, not throughput, latency, power efficiency or broad software-ecosystem parity.
- facts: [] → [{"label":"MiniCPM5-2B deployment","value":"Kunlunxin and FlagOS report Day-0 adaptation and inference validation, with a BF16 deployment image published through ModelScope and Hugging Face.","evidence":"vendor","as_of":"2026-09-10","source_ids":["kunlunxin-minicpm-2026-09-10"]}]
- source_ids: ["kunlunxin"] → ["kunlunxin","kunlunxin-minicpm-2026-09-10"]
- reviewed_at: 2026-09-11 → 2026-09-16
Updated: MetaX
4 field changes
- assessment: The baseline does not establish frontier performance parity. → MetaX reports broadening model and framework coverage, including ATRIA Dawn Preview support on release day. The current evidence is vendor-reported compatibility and operator coverage; it does not establish application performance, correctness across workloads or frontier parity.
- facts: [] → [{"label":"Software compatibility coverage","value":"MetaX reports 37 flagship-model Day-0 adaptations since December 2025 and full support for PyTorch 2.8's 2,410 GPU operators; no reproducible workload benchmark accompanied the claim.","evidence":"vendor","as_of":"2026-09-14","source_ids":["metax-atria-2026-09-14"]}]
- source_ids: ["hbm-reuters"] → ["hbm-reuters","metax-atria-2026-09-14"]
- reviewed_at: 2026-09-11 → 2026-09-16
Updated: Moore Threads
8 field changes
- assessment: Compatibility claims need workload-level verification. → Moore Threads now has company-reported commercial evidence from a delivered RMB 660 million KUAE cluster contract, while JD Cloud's 100,000-GPU cluster remains an announced plan rather than an operating deployment. The disclosed utilization and scaling figures are self-measured and do not provide a matched current foreign-hardware comparison.
- key_fact.label: Current platform → Delivered cluster contract
- key_fact.value: MTT S5000 and MUSA → Company reports a RMB 660 million KUAE contract fully delivered and recognized in first-half revenue
- key_fact.as_of: 2026-09-11 → 2026-06-30
- key_fact.source_ids: ["mthreads"] → ["moore-threads-2026-h1-report"]
- facts: [] → [{"label":"First-half commercial scale","value":"Company reports RMB 1.736 billion of first-half revenue, up 147.42% year over year, and an attributable net loss of RMB 11.56 million.","evidence":"vendor","as_of":"2026-06-30","source_ids":["moore-threads-2026-h1-report"]},{"label":"Vendor-reported KUAE training metrics","value":"Company reports Dense-model MFU up to 60%, MoE-model MFU up to 40%, 95% linear scaling efficiency and effective training time above 90%; workload, power and matched comparator details were not disclosed in the filing.","evidence":"vendor","as_of":"2026-06-30","source_ids":["moore-threads-2026-h1-report"]},{"label":"JD Cloud cluster plan","value":"JD Cloud announced a plan to build a 100,000-GPU cluster based on Moore Threads; no delivery schedule, completed card count, topology or operational results were disclosed.","evidence":"vendor","as_of":"2026-09-09","source_ids":["moore-threads-jd-cloud-2026-09-09"]}]
- source_ids: ["mthreads"] → ["mthreads","moore-threads-2026-h1-report","moore-threads-jd-cloud-2026-09-09"]
- reviewed_at: 2026-09-11 → 2026-09-16
Added: Kunlunxin and FlagOS publish a MiniCPM5-2B deployment image
Added: MetaX reports same-day adaptation of ATRIA Dawn Preview
Added: Moore Threads reports first-half revenue scale and a delivered KUAE cluster contract
Added: JD Cloud announces planned 100,000-GPU cluster based on Moore Threads
Updated: data/research-state.yaml
4 field changes
- last_research_at: 2026-09-15 → 2026-09-16
- last_automated_run_at: 2026-09-15 → 2026-09-16
- coverage: The September 15 daily review screened all 12 priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Three previously omitted August half-year filings produced material production, qualification and delivery-scale backfills for AMEC, NAURA and JCET. No new qualifying September 15 accelerator, memory, optical or benchmark event was verified. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts; only the five profiles whose primary pages or filings were actually reviewed advanced. → The September 16 daily review screened all 12 priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Current primary disclosures produced software-compatibility updates for Kunlunxin and MetaX, a planned 100,000-GPU JD Cloud cluster for Moore Threads, and a late-added Moore Threads half-year filing with commercial-scale evidence. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts; only the three profiles whose current primary pages or filing were actually reviewed advanced.
- runs: [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]}] → [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-16","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","kunlunxin","metax","moore-threads"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 9-16. Added Moore Threads' September 9 primary announcement that JD Cloud plans a 100,000-card full-function GPU cluster. The record treats this as a plan rather than delivery evidence because no schedule, GPU model, completed card count, topology, interconnect, power, recovery or workload result was disclosed. Backfilled Moore Threads' August half-year report: company-reported revenue was RMB 1.736 billion, attributable net loss was RMB 11.56 million, and a RMB 660 million KUAE sales contract was fully delivered and recognized. Its Dense and MoE MFU, scaling-efficiency and effective- training-time figures remain vendor measurements without a matched workload, power or current foreign-hardware comparison. Added Kunlunxin's downloadable FlagOS MiniCPM5-2B deployment image and MetaX's ATRIA Dawn Preview Day-0 adaptation plus reported PyTorch operator coverage as compatibility evidence, not performance or CUDA parity. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current coverage. No matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was verified. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the half-year filing is explicitly late-added.","source_ids":["moore-threads-2026-h1-report","moore-threads-jd-cloud-2026-09-09","kunlunxin-minicpm-2026-09-10","metax-atria-2026-09-14","mthreads","kunlunxin","mlperf-training","mlperf-inference","inferencex"]}]
Updated: Continuous inference benchmark dashboard
1 field changes
- reviewed_at: 2026-09-15 → 2026-09-16
Added: Kunlunxin and FlagOS complete MiniCPM5-2B deployment adaptation
Updated: AI accelerator portfolio
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-16
Added: MetaX reports ATRIA Dawn Preview Day-0 adaptation
Updated: MLPerf Inference v6.0 submissions
1 field changes
- reviewed_at: 2026-09-15 → 2026-09-16
Updated: MLPerf Training v6.0 submissions
1 field changes
- reviewed_at: 2026-09-15 → 2026-09-16
Added: 2026 half-year report
Added: JD Cloud and Moore Threads announce planned 100,000-GPU cluster
Updated: GPU and MUSA portfolio
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-16

2026-09-15d3a566e ↗

Backfill H1 equipment and packaging evidence

Updated: data/benchmarks.yaml
2 field changes
- assessment: Follow-up through September 14 did not verify a matched independent comparison of current Chinese accelerators against GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability. → Follow-up through September 15 did not verify a matched independent comparison of current Chinese accelerators against GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability.
- reviewed_at: 2026-09-14 → 2026-09-15
Updated: AMEC
8 field changes
- assessment: Company-level tool capabilities are not proof of complete process independence. → AMEC's half-year report provides production-line, customer-validation and repeat-order evidence across etch and tungsten deposition. It does not name customers or disclose advanced-node yield, uptime, throughput, installed capacity or complete-process independence.
- key_fact.label: Strategic role → Etch deployment scale
- key_fact.value: Etch equipment for semiconductor manufacturing → Company reports three ICP etch product families in volume production on lines at more than 50 customers
- key_fact.as_of: 2026-09-11 → 2026-06-30
- key_fact.source_ids: ["amec"] → ["amec-2026-h1-report"]
- facts: [{"label":"CSEAC 2026 tool launches","value":"Six announced tools span very-high-aspect-ratio etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; customer qualification remains unverified.","evidence":"reported","as_of":"2026-09-01","source_ids":["amec-cseac-2026-09-05"]}] → [{"label":"Advanced etch production","value":"Company reports Primo Nanova LUX-Cryo in stable volume production on a customer's next-generation product line, while Primo Angnova remained in customer certification with multiple repeat orders.","evidence":"vendor","as_of":"2026-06-30","source_ids":["amec-2026-h1-report"]},{"label":"Tungsten deposition qualification","value":"Company reports its tungsten CVD, high-aspect-ratio CVD and ALD series passed onsite validation at key memory and logic customers and received repeat production orders; customers and production metrics were not disclosed.","evidence":"vendor","as_of":"2026-06-30","source_ids":["amec-2026-h1-report"]},{"label":"CSEAC 2026 tool launches","value":"Six announced tools span very-high-aspect-ratio etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; customer qualification remains unverified.","evidence":"reported","as_of":"2026-09-01","source_ids":["amec-cseac-2026-09-05"]}]
- source_ids: ["amec","amec-cseac-2026-09-05"] → ["amec","amec-2026-h1-report","amec-cseac-2026-09-05"]
- reviewed_at: 2026-09-13 → 2026-09-15
Updated: Iluvatar CoreX
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-15
Updated: JCET
9 field changes
- assessment: Broad packaging strength does not establish frontier CoWoS-equivalent capacity. → JCET's half-year report provides company evidence of XDFOI volume production and customer validation of silicon-photonics engine samples. It does not establish frontier compute-plus-HBM package yield, capacity, thermal performance or domestic input provenance.
- key_fact.label: Technology → XDFOI production status
- key_fact.value: XDFOI heterogeneous integration research and platforms → Company reports that its high-density heterogeneous-integration platform reached volume production
- key_fact.evidence: primary → vendor
- key_fact.as_of: 2026-09-11 → 2026-06-30
- key_fact.source_ids: ["jcet"] → ["jcet-2026-h1-report"]
- facts: [] → [{"label":"Computing-electronics mix","value":"First-half computing-electronics revenue grew 40.4% year over year and represented 30.1% of company revenue.","evidence":"vendor","as_of":"2026-06-30","source_ids":["jcet-2026-h1-report"]},{"label":"Silicon-photonics validation","value":"Company reports delivering XDFOI-based silicon-photonics engine samples that passed customer validation and working with multiple customers on packaging, thermal and reliability optimization; volume and capacity were not disclosed.","evidence":"vendor","as_of":"2026-06-30","source_ids":["jcet-2026-h1-report"]}]
- source_ids: ["jcet"] → ["jcet","jcet-2026-h1-report"]
- reviewed_at: 2026-09-11 → 2026-09-15
Updated: Kingsemi
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-15
Updated: NAURA
8 field changes
- assessment: A qualified tool covers a process step, not an entire sovereign fab. → NAURA reports broad commercial scale and more than 1,000 cumulative deliveries in each of three tool categories. Delivery counts do not establish advanced-node customer qualification, production yield or domestic provenance for critical components, and a qualified tool still covers only one process step rather than a sovereign fab flow.
- key_fact.label: Portfolio → Deployment scale
- key_fact.value: Etch, PVD, CVD, wet processing and more → Company reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnace systems
- key_fact.as_of: 2026-09-11 → 2026-06-30
- key_fact.source_ids: ["naura"] → ["naura-2026-h1-report"]
- facts: [] → [{"label":"First-half commercial scale","value":"Company reported RMB 20.161 billion of first-half revenue, up 24.90% year over year, and RMB 2.677 billion of R&D expense, up 28.87%.","evidence":"vendor","as_of":"2026-06-30","source_ids":["naura-2026-h1-report"]},{"label":"Advanced packaging expansion","value":"Company reports that hybrid-bonding and high-end TSV electroplating tools were rolled out; the filing does not disclose named-customer acceptance, yield or production volume.","evidence":"vendor","as_of":"2026-06-30","source_ids":["naura-2026-h1-report"]}]
- source_ids: ["naura"] → ["naura","naura-2026-h1-report"]
- reviewed_at: 2026-09-11 → 2026-09-15
Added: AMEC reports production deployment and repeat orders across etch and tungsten deposition
Added: JCET reports XDFOI volume production and silicon-photonics customer validation
Added: NAURA reports thousand-unit delivery scale in three semiconductor-equipment categories
Updated: data/research-state.yaml
4 field changes
- last_research_at: 2026-09-14 → 2026-09-15
- last_automated_run_at: 2026-09-14 → 2026-09-15
- coverage: The September 7–13 complete weekly edition reconstructs four pre-baseline days and covers the three days from the September 11 baseline through period close using ten canonical events, selected primary pages and reporting, and overlapping English and Chinese gap searches. It adds one Biren software-stack event. Coverage is incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts; only Biren's profile review date advanced in this briefing run. → The September 15 daily review screened all 12 priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Three previously omitted August half-year filings produced material production, qualification and delivery-scale backfills for AMEC, NAURA and JCET. No new qualifying September 15 accelerator, memory, optical or benchmark event was verified. Coverage remains incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts; only the five profiles whose primary pages or filings were actually reviewed advanced.
- runs: [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]}] → [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-15","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","iluvatar","jcet","kingsemi"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with overlapping English and Chinese searches covering September 8-15. No new qualifying in-window accelerator, memory, optical-interconnect or semiconductor-equipment event was verified. A filing-gap review found three material August disclosures that were missing from the canonical record. AMEC's company report says three ICP etch families were in volume production on lines at more than 50 customers, with stable production use or repeat orders for additional advanced etch systems and repeat production orders after onsite validation of tungsten deposition systems at key memory and logic customers. NAURA reports more than 1,000 cumulative deliveries each for etch, PVD and vertical furnaces, RMB 20.161 billion of first-half revenue, and rollout of hybrid-bonding and high-end TSV electroplating tools. JCET reports XDFOI volume production, customer validation of silicon-photonics engine samples, and 40.4% year-over-year growth in computing-electronics revenue. Each claim remains vendor evidence: unnamed customers, unspecified process nodes, and undisclosed yield, capacity, uptime, throughput or production volume limit the conclusions. Reviewed Iluvatar and Kingsemi's current official pages without accepting a new production or benchmark claim. Rechecked MLPerf Training and Inference v6.0 repository heads and InferenceX's current public coverage; no matched contemporary Chinese result against current GB300, MI355X or TPU v7-class systems was found. Search-only priority coverage did not advance review dates. Historical reports, corrections, forecasts and evidence labels were preserved; the August events are explicitly late-added rather than dated as September events.","source_ids":["amec-2026-h1-report","naura-2026-h1-report","jcet-2026-h1-report","iluvatar","kingsemi","mlperf-training","mlperf-inference","inferencex"]}]
Added: 2026 half-year report
Updated: Compute platform portfolio
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-15
Updated: Continuous inference benchmark dashboard
1 field changes
- reviewed_at: 2026-09-14 → 2026-09-15
Added: 2026 half-year report
Updated: Coating and developing equipment
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-15
Updated: MLPerf Inference v6.0 submissions
1 field changes
- reviewed_at: 2026-09-14 → 2026-09-15
Updated: MLPerf Training v6.0 submissions
1 field changes
- reviewed_at: 2026-09-14 → 2026-09-15
Added: 2026 half-year report

2026-09-14f0e2694 ↗

Publish complete September 7 weekly briefing

Added: Modules and software advance; memory and qualification still bind
Updated: Biren
4 field changes
- facts: [{"label":"First-half 2026 commercialization","value":"Company reports RMB 1.236 billion revenue (up 1,997.6% year on year) and a RMB 377 million loss for the period in its September 7 half-year summary.","evidence":"vendor","as_of":"2026-06-30","source_ids":["biren-2026-h1"]},{"label":"Model compatibility","value":"Reports functional and accuracy validation of GLM-5.3-Flash on Bili 166M using SGLang and BIRENSUPA; no matched hardware speedup established.","evidence":"vendor","as_of":"2026-09-07","source_ids":["biren-glm-2026-09-07"]},{"label":"Optical-interconnect research","value":"September 2 Shanghai University laboratory agreement includes planned AI-chip optical-interconnect and cooling research, not production qualification.","evidence":"vendor","as_of":"2026-09-02","source_ids":["biren-shanghai-lab-2026-09-10"]}] → [{"label":"First-half 2026 commercialization","value":"Company reports RMB 1.236 billion revenue (up 1,997.6% year on year) and a RMB 377 million loss for the period in its September 7 half-year summary.","evidence":"vendor","as_of":"2026-06-30","source_ids":["biren-2026-h1"]},{"label":"Model compatibility","value":"Reports functional and accuracy validation of GLM-5.3-Flash on Bili 166M using SGLang and BIRENSUPA; no matched hardware speedup established.","evidence":"vendor","as_of":"2026-09-07","source_ids":["biren-glm-2026-09-07"]},{"label":"PyTorch backend","value":"Company says it released TorchSUPA, a PrivateUse1 backend plugin with common API, collective-communication, graph-capture and profiling support; no public code repository was identified in this review.","evidence":"vendor","as_of":"2026-09-07","source_ids":["biren-torchsupa-2026-09-07"]},{"label":"Optical-interconnect research","value":"September 2 Shanghai University laboratory agreement includes planned AI-chip optical-interconnect and cooling research, not production qualification.","evidence":"vendor","as_of":"2026-09-02","source_ids":["biren-shanghai-lab-2026-09-10"]}]
- source_ids: ["biren","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10"] → ["biren","biren-2026-h1","biren-glm-2026-09-07","biren-torchsupa-2026-09-07","biren-shanghai-lab-2026-09-10"]
- reviewed_at: 2026-09-12 → 2026-09-14
- layers: ["accelerators"] → ["accelerators","software"]
Added: Biren announces TorchSUPA PyTorch backend plugin
Updated: data/research-state.yaml
2 field changes
- coverage: September 14 sweep used September 7–14 overlapping English and Chinese searches across all 12 priority-1 companies plus Beijing Guowang, H3C and Hwatsing, the next three oldest-reviewed non-priority profiles. H3C and Hwatsing gained substantive evidence; Guowang's patent record was rechecked without a verified material change. Other priority review dates were not advanced for search-only coverage. This is an overlap sweep, not a full revalidation of the 31-company directory. → The September 7–13 complete weekly edition reconstructs four pre-baseline days and covers the three days from the September 11 baseline through period close using ten canonical events, selected primary pages and reporting, and overlapping English and Chinese gap searches. It adds one Biren software-stack event. Coverage is incomplete for private deployments, unindexed Chinese disclosures, tenders and individual benchmark artifacts; only Biren's profile review date advanced in this briefing run.
- runs: [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]}] → [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","biren","cambricon","cxmt","enflame","eoptolink","huawei","hwatsing","iluvatar","innolight","metax","naura","rslaser","smic","ymtc","yuliangsheng","zhiqi"],"summary":"Ran the weekly helper for the previous fully closed Monday-Sunday period and retained the September 11 period-to-date edition unchanged. Audited the ten selected in-period events and ran overlapping English and Chinese gap searches across accelerators, software, memory, optical interconnects, lithography and packaging equipment. Added Biren's September 7 vendor account of an August 28 ecosystem meetup as a canonical source and event: it says TorchSUPA uses PyTorch PrivateUse1 and supports common APIs, collectives and FSDP2, graph capture and profiling. No public code repository was identified, so the record does not infer operator coverage, maintainability, upstream acceptance, performance or CUDA parity. The complete briefing preserves all historical reports and states that no dated forecast moved. It uses the current September 14 benchmark review: no matched contemporary Chinese result against GB300, MI355X or TPU v7 was verified in MLPerf v6.0 or InferenceX coverage. Because systematic coverage began September 11, the September 7-13 reconstruction remains incomplete for private deployments, unindexed disclosures, tenders and individual benchmark artifacts. Search-only coverage did not advance other company review dates.","source_ids":["biren-torchsupa-2026-09-07","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","hwatsing-panel-cmp-2026-09-09","duv-ft","mlperf-training","mlperf-inference","inferencex"]}]
Added: Biren announces TorchSUPA PyTorch backend at ecosystem meetup

2026-09-1491b7a3f ↗

Research September 14: interconnect and packaging equipment

Updated: data/benchmarks.yaml
2 field changes
- assessment: Follow-up through September 13 did not verify a matched independent comparison of current Chinese accelerators against GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability. → Follow-up through September 14 did not verify a matched independent comparison of current Chinese accelerators against GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability.
- reviewed_at: 2026-09-13 → 2026-09-14
Updated: Beijing Guowang
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-14
Updated: H3C
9 field changes
- assessment: A domestic network-system brand does not establish domestic provenance for all silicon. → Open interconnect work and a reported customer deployment are meaningful integration signals, but do not establish domestic provenance for all silicon or performance parity with InfiniBand.
- key_fact.label: Strategic role → Deployment evidence
- key_fact.value: Ethernet infrastructure for AI clusters → H3C reports a deployed DDC lossless-network project with Xiaohongshu and participation in G-Link and ODCC standards work; no matched InfiniBand workload result was disclosed.
- key_fact.evidence: vendor → primary
- key_fact.as_of: 2026-09-11 → 2026-09-04
- key_fact.source_ids: ["h3c"] → ["h3c-odx-2026-09-04"]
- facts: [] → [{"label":"Open scale-up direction","value":"H3C describes UXM module support, G-Link development and NPO/CPO interfaces with 112G/224G compatibility and a reserved 448G SerDes path.","evidence":"primary","as_of":"2026-09-04","source_ids":["h3c-odx-2026-09-04"]}]
- source_ids: ["h3c"] → ["h3c","h3c-odx-2026-09-04"]
- reviewed_at: 2026-09-11 → 2026-09-14
Updated: Hwatsing
9 field changes
- assessment: Polishing and thinning quality are essential to downstream packaging yield. → Polishing and thinning quality are essential to downstream packaging yield. Delivery to a production line does not by itself disclose customer acceptance, yield, uptime or volume.
- key_fact.label: Product portfolio → Panel-level CMP delivery
- key_fact.value: 12-inch CMP and advanced-packaging equipment → Master-P510APEX was reported shipped to an advanced-packaging production line for use.
- key_fact.evidence: vendor → reported
- key_fact.as_of: 2026-09-11 → 2026-09-09
- key_fact.source_ids: ["hwatsing"] → ["hwatsing-panel-cmp-2026-09-09"]
- facts: [] → [{"label":"Process scope","value":"The vendor describes automated endpoint detection and multiple polishing modes for large-format panel-level packaging; customer and production metrics were not disclosed.","evidence":"reported","as_of":"2026-09-09","source_ids":["hwatsing-panel-cmp-2026-09-09"]}]
- source_ids: ["hwatsing"] → ["hwatsing","hwatsing-panel-cmp-2026-09-09"]
- reviewed_at: 2026-09-11 → 2026-09-14
Added: H3C presents open interconnect work and an awarded DDC deployment
Added: Hwatsing ships panel-level CMP equipment to an advanced-packaging line
Updated: data/research-state.yaml
4 field changes
- last_research_at: 2026-09-13 → 2026-09-14
- last_automated_run_at: 2026-09-13 → 2026-09-14
- coverage: September 13 sweep used September 6–13 overlapping English and Chinese searches across all 12 priority-1 companies plus ACM Research Shanghai, CIOMP and Empyrean, the first three alphabetically among the oldest-reviewed non-priority profiles. AMEC, Cambricon and Huawei gained substantive evidence; the three oldest profiles were rechecked against current company pages without a verified material change. Other priority review dates were not advanced for search-only coverage. This is an overlap sweep, not a full revalidation of the 31-company directory. → September 14 sweep used September 7–14 overlapping English and Chinese searches across all 12 priority-1 companies plus Beijing Guowang, H3C and Hwatsing, the next three oldest-reviewed non-priority profiles. H3C and Hwatsing gained substantive evidence; Guowang's patent record was rechecked without a verified material change. Other priority review dates were not advanced for search-only coverage. This is an overlap sweep, not a full revalidation of the 31-company directory.
- runs: [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]}] → [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-14","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","guowang","h3c","hwatsing"],"summary":"Screened all priority-1 profiles and the next three oldest-reviewed non-priority profiles with September 7–14 overlapping English and Chinese searches. Added H3C primary evidence for its reported Xiaohongshu DDC deployment, G-Link and ODCC participation, UXM module support and open optical-interconnect direction. The disclosure did not provide a matched InfiniBand workload benchmark, cluster topology, switch-silicon provenance or independent customer data. Added reporting sourced to Hwatsing's company announcement that its Master-P510APEX panel-level CMP system was shipped to an advanced-packaging production line; kept shipment distinct from customer acceptance, yield, uptime, throughput and volume evidence. Reopened Guowang's patent record and the current H3C and Hwatsing company pages; Guowang gained no new production evidence. Search-only coverage of the other priority profiles did not advance their review dates. Rechecked MLPerf Training and Inference v6.0 repository heads and submitter directories plus InferenceX's current public coverage. The MLPerf repositories had no newer result push and no current Chinese accelerator submitter was found; InferenceX listed current GB300, MI355X and TPU v7 systems but no equivalent current Chinese hardware result. No matched latest-versus-latest parity claim was added. Historical reports, corrections, forecasts, evidence labels and prior run records were preserved.","source_ids":["h3c-odx-2026-09-04","hwatsing-panel-cmp-2026-09-09","guowang","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]}]
Updated: Optical-system patent
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-14
Added: H3C presents open AI-infrastructure and DDC deployment work at ODX 2026
Updated: Data-center networking portfolio
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-14
Added: Hwatsing panel-level CMP system enters a production line
Updated: CMP and related process equipment
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-14
Updated: Continuous inference benchmark dashboard
1 field changes
- reviewed_at: 2026-09-13 → 2026-09-14
Updated: MLPerf Inference v6.0 submissions
1 field changes
- reviewed_at: 2026-09-13 → 2026-09-14
Updated: MLPerf Training v6.0 submissions
1 field changes
- reviewed_at: 2026-09-13 → 2026-09-14

2026-09-137929119 ↗

Research September 13: accelerator software and AMEC tools

Updated: data/benchmarks.yaml
2 field changes
- assessment: The baseline and September 12 scoped follow-up did not verify a matched independent comparison of current Chinese accelerators against GB300, MI355X, or TPU v7. The follow-up checked MLPerf v6.0 repository listings and InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability. → Follow-up through September 13 did not verify a matched independent comparison of current Chinese accelerators against GB300, MI355X, or TPU v7. The review checked MLPerf v6.0 repository heads and submitter listings plus InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability.
- reviewed_at: 2026-09-12 → 2026-09-13
Updated: ACM Research Shanghai
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-13
Updated: AMEC
3 field changes
- facts: [] → [{"label":"CSEAC 2026 tool launches","value":"Six announced tools span very-high-aspect-ratio etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; customer qualification remains unverified.","evidence":"reported","as_of":"2026-09-01","source_ids":["amec-cseac-2026-09-05"]}]
- source_ids: ["amec"] → ["amec","amec-cseac-2026-09-05"]
- reviewed_at: 2026-09-11 → 2026-09-13
Updated: Cambricon
3 field changes
- facts: [] → [{"label":"Upstream software participation","value":"PyTorch Foundation Platinum member with reported contributions spanning torch.compile, runtime, distributed computing, AMP, data loading and profiling.","evidence":"primary","as_of":"2026-09-08","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09"]}]
- source_ids: ["hbm-reuters","cambricon"] → ["hbm-reuters","cambricon","pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09"]
- reviewed_at: 2026-09-11 → 2026-09-13
Updated: CIOMP
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-13
Updated: Empyrean
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-13
Updated: Huawei
3 field changes
- facts: [{"label":"Atlas 350 / 950PR","value":"112 GB HBM; 1.4 TB/s; 425 TFLOPS BF16; up to 600 W","evidence":"vendor","as_of":"2026-09-11","source_ids":["atlas350"]}] → [{"label":"PyTorch accelerator integration","value":"Huawei is a PyTorch Foundation Platinum member and co-chairs its Accelerator Integration Working Group; governance participation does not prove workload parity.","evidence":"primary","as_of":"2026-09-09","source_ids":["pytorch-china-2026-09-09"]},{"label":"Atlas 350 / 950PR","value":"112 GB HBM; 1.4 TB/s; 425 TFLOPS BF16; up to 600 W","evidence":"vendor","as_of":"2026-09-11","source_ids":["atlas350"]}]
- source_ids: ["ascend","hbm-reuters"] → ["ascend","hbm-reuters","pytorch-china-2026-09-09"]
- reviewed_at: 2026-09-11 → 2026-09-13
Added: AMEC expands its process-tool portfolio with six CSEAC launches
Added: Cambricon joins PyTorch Foundation governance as accelerator integration broadens
Updated: data/research-state.yaml
5 field changes
- last_research_at: 2026-09-12 → 2026-09-13
- last_automated_run_at: 2026-09-12 → 2026-09-13
- last_run_status: partial → updated
- coverage: September 12 sweep used September 4–12 overlapping English and Chinese searches across all 12 priority-1 companies plus ACM Research Shanghai, Biren and CIOMP, selected alphabetically from the oldest-reviewed remaining profiles (all tied on September 11). Substantive source review updated four profiles; other profile review dates were not advanced for search-only coverage. Original-source access and some follow-up verification remain incomplete; this is not a full revalidation of the 31-company directory. → September 13 sweep used September 6–13 overlapping English and Chinese searches across all 12 priority-1 companies plus ACM Research Shanghai, CIOMP and Empyrean, the first three alphabetically among the oldest-reviewed non-priority profiles. AMEC, Cambricon and Huawei gained substantive evidence; the three oldest profiles were rechecked against current company pages without a verified material change. Other priority review dates were not advanced for search-only coverage. This is an overlap sweep, not a full revalidation of the 31-company directory.
- runs: [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]}] → [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]},{"date":"2026-09-13","status":"updated","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","ciomp","empyrean"],"summary":"Screened all priority-1 profiles and the three oldest-reviewed non-priority profiles with overlapping English and Chinese searches. Added PyTorch Foundation primary evidence that Cambricon joined as a Platinum member with Governing Board and Technical Advisory Council representation and reported upstream work across compile, runtime and distributed areas. Added Huawei's existing Platinum membership and co-chair role in the Foundation's Accelerator Integration Working Group. These are ecosystem and governance signals, not performance results. Closed the prior AMEC follow-up using detailed Science and Technology Daily reporting on six CSEAC launches across etch, PECVD, ALD, metal-film deposition and silicon-carbide epitaxy; kept specifications distinct from customer qualification and volume-production evidence. Reviewed current official portfolio and news pages for ACM Research Shanghai, CIOMP and Empyrean; no new qualifying production, customer or complete-flow milestone was verified, so their facts were unchanged while their actual review dates advanced. Rechecked the Financial Times DUV story, HBM pricing coverage, the September 11 optics report and Enflame's IPO follow-up during the overlap window; no additional claim was accepted. Reviewed the FCC's final component-rule coverage but did not attach an entity-specific restriction to InnoLight or Eoptolink because the rule did not name them. MLPerf Training and Inference v6.0 repository heads were unchanged, and the closed inference submitter list still showed no matched current Chinese accelerator. InferenceX coverage still did not supply an equivalent Chinese result. No independent latest-versus-latest parity claim was added. Historical news, reports, forecasts, evidence labels and prior run records were preserved.","source_ids":["pytorch-cambricon-2026-09-07","pytorch-china-2026-09-09","amec-cseac-2026-09-05","acm","ciomp","empyrean","duv-ft","hbm-reuters","cioe-optics-2026-09-11","enflame-ipo","mlperf-training","mlperf-inference","inferencex"]}]
Added: AMEC launches six self-developed process tools at CSEAC 2026
Updated: EDA portfolio
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-13
Updated: Continuous inference benchmark dashboard
1 field changes
- reviewed_at: 2026-09-12 → 2026-09-13
Updated: MLPerf Inference v6.0 submissions
1 field changes
- reviewed_at: 2026-09-12 → 2026-09-13
Updated: MLPerf Training v6.0 submissions
1 field changes
- reviewed_at: 2026-09-12 → 2026-09-13
Added: Cambricon joins the PyTorch Foundation as a Platinum member
Added: PyTorch Conference China 2026 advances the open-source AI stack

2026-09-12304c5c4 ↗

Research September 12: optics shipments and Biren evidence

Updated: data/benchmarks.yaml
3 field changes
- assessment: The baseline search did not find a matched independent comparison of the latest Chinese accelerators against GB300, MI355X, or TPU v7. Absence of a public result is not evidence of incapability. → The baseline and September 12 scoped follow-up did not verify a matched independent comparison of current Chinese accelerators against GB300, MI355X, or TPU v7. The follow-up checked MLPerf v6.0 repository listings and InferenceX's published coverage, not every result artifact. Absence of a verified public result is not evidence of incapability.
- reviewed_at: 2026-09-11 → 2026-09-12
- studies: [{"title":"MLPerf Training / Inference v6.0","evidence":"standardized","date":"2026","finding":"Standardized submitter-run results; no matched latest Chinese-versus-foreign comparison established in this baseline.","source_ids":["mlperf-training","mlperf-inference"]},{"title":"InferenceX","evidence":"measured","date":"Continuous","finding":"Reproducible independent inference runs. GB300, MI355X and TPU v7 are represented; equivalent current Chinese results were not verified.","source_ids":["inferencex"]},{"title":"Ascend operational field study","evidence":"measured","date":"2026-07","finding":"A small independent preprint documents deployment friction on older Ascend 910 devices. No matched NVIDIA control.","source_ids":["ascend-field"]},{"title":"HiFloat4 pretraining","evidence":"vendor","date":"2026-04","finding":"Demonstrates training progress on dense and MoE models; does not measure frontier time-to-quality against foreign hardware.","source_ids":["hifloat4"]}] → [{"title":"Biren GLM-5.3-Flash functional validation","evidence":"vendor","date":"2026-09-07","finding":"Vendor reports functional and accuracy checks on Bili 166M with SGLang and BIRENSUPA. Compatibility evidence only; no matched foreign-hardware throughput, latency or power comparison.","source_ids":["biren-glm-2026-09-07"]},{"title":"MLPerf Training / Inference v6.0","evidence":"standardized","date":"2026","finding":"Standardized submitter-run results; no matched latest Chinese-versus-foreign comparison established in this baseline.","source_ids":["mlperf-training","mlperf-inference"]},{"title":"InferenceX","evidence":"measured","date":"Continuous","finding":"Reproducible independent inference runs. GB300, MI355X and TPU v7 are represented; equivalent current Chinese results were not verified.","source_ids":["inferencex"]},{"title":"Ascend operational field study","evidence":"measured","date":"2026-07","finding":"A small independent preprint documents deployment friction on older Ascend 910 devices. No matched NVIDIA control.","source_ids":["ascend-field"]},{"title":"HiFloat4 pretraining","evidence":"vendor","date":"2026-04","finding":"Demonstrates training progress on dense and MoE models; does not measure frontier time-to-quality against foreign hardware.","source_ids":["hifloat4"]}]
Updated: Biren
3 field changes
- facts: [] → [{"label":"First-half 2026 commercialization","value":"Company reports RMB 1.236 billion revenue (up 1,997.6% year on year) and a RMB 377 million loss for the period in its September 7 half-year summary.","evidence":"vendor","as_of":"2026-06-30","source_ids":["biren-2026-h1"]},{"label":"Model compatibility","value":"Reports functional and accuracy validation of GLM-5.3-Flash on Bili 166M using SGLang and BIRENSUPA; no matched hardware speedup established.","evidence":"vendor","as_of":"2026-09-07","source_ids":["biren-glm-2026-09-07"]},{"label":"Optical-interconnect research","value":"September 2 Shanghai University laboratory agreement includes planned AI-chip optical-interconnect and cooling research, not production qualification.","evidence":"vendor","as_of":"2026-09-02","source_ids":["biren-shanghai-lab-2026-09-10"]}]
- source_ids: ["biren"] → ["biren","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10"]
- reviewed_at: 2026-09-11 → 2026-09-12
Updated: Enflame
2 field changes
- facts: [{"label":"S60 compute chiplet","value":"TSMC N6, identified by independent teardown","evidence":"measured","as_of":"2026-02-03","source_ids":["enflame-teardown"]}] → [{"label":"IPO-funded development plans","value":"Reuters reports IPO proceeds are intended mainly for fifth- and sixth-generation AI chips, software and large computing systems; funding plans are not delivered products.","evidence":"reported","as_of":"2026-09-11","source_ids":["enflame-ipo"]},{"label":"S60 compute chiplet","value":"TSMC N6, identified by independent teardown","evidence":"measured","as_of":"2026-02-03","source_ids":["enflame-teardown"]}]
- reviewed_at: 2026-09-11 → 2026-09-12
Updated: Eoptolink
3 field changes
- facts: [] → [{"label":"1.6T shipment progression","value":"Company reports higher second-quarter than first-quarter 2026 shipments; faster second-half growth is an expectation, not a realized result.","evidence":"reported","as_of":"2026-09-11","source_ids":["cioe-optics-2026-09-11"]}]
- source_ids: ["eop-dsp","eoptolink"] → ["eop-dsp","eoptolink","cioe-optics-2026-09-11"]
- reviewed_at: 2026-09-11 → 2026-09-12
Updated: InnoLight
3 field changes
- facts: [] → [{"label":"1.6T commercialization","value":"Company says silicon-photonics modules are scaling with shipments rising each quarter, according to National Business Daily's September 11 reporting.","evidence":"reported","as_of":"2026-09-11","source_ids":["cioe-optics-2026-09-11"]}]
- source_ids: ["innolight"] → ["innolight","cioe-optics-2026-09-11"]
- reviewed_at: 2026-09-11 → 2026-09-12
Added: Biren's half-year summary reports revenue growth alongside continuing losses
Added: Biren reports GLM-5.3-Flash inference adaptation on Bili 166M
Added: Biren and Shanghai University establish a joint intelligent-chip laboratory
Added: CIOE reporting distinguishes optical-module shipment ramps from next-generation samples
Updated: data/research-state.yaml
5 field changes
- last_research_at: 2026-09-11 → 2026-09-12
- last_automated_run_at: — → 2026-09-12
- last_run_status: baseline → partial
- coverage: Sourced baseline from the initial research; per-company review dates are shown. → September 12 sweep used September 4–12 overlapping English and Chinese searches across all 12 priority-1 companies plus ACM Research Shanghai, Biren and CIOMP, selected alphabetically from the oldest-reviewed remaining profiles (all tied on September 11). Substantive source review updated four profiles; other profile review dates were not advanced for search-only coverage. Original-source access and some follow-up verification remain incomplete; this is not a full revalidation of the 31-company directory.
- runs: [] → [{"date":"2026-09-12","status":"partial","companies":["aishengna","amec","cambricon","cxmt","enflame","eoptolink","huawei","innolight","naura","smic","ymtc","yuliangsheng","acm","biren","ciomp"],"summary":"These 15 companies were screened, not all fully revalidated. Read original CIOE reporting and optical suppliers' public sites; added reported shipment progression for InnoLight and Eoptolink. Read Biren's dated primary announcements; added software-validation and research-laboratory events, preserving the distinction between publication and event dates. Reviewed the indexed original September 11 Reuters Enflame report and added planned use of IPO proceeds without duplicating its existing IPO event. Only these four company review dates advanced. Priority-company searches also revisited the September 10 HBM pricing report and July DUV reporting; no additional chip-performance or qualified-tool claim was accepted. AMEC's official site and the full Financial Times DUV article could not be retrieved. Follow up AMEC's six-tool announcement, the DUV testing chronology, reported optical-module export restrictions, and memory/equipment filings; do not infer qualification from headlines. Downloaded and visually checked Biren's primary half-year infographic; added its revenue, loss, R&D and unnamed-customer delivery claims with vendor attribution, not an independent audit. ACM and CIOMP searches did not establish a verified new qualifying milestone; their profile dates remain unchanged. Checked MLPerf Training v6.0 repository listings, Inference v6.0 closed-submitter listings via GitHub, and InferenceX's public coverage. No matched current China-versus-frontier result was verified; individual result artifacts were not exhaustively audited. Historical news, period briefings, forecasts and accelerator specifications are retained. Research scope is partial independently of GitHub sync or Cloudflare publication status.","source_ids":["cioe-optics-2026-09-11","biren-2026-h1","biren-glm-2026-09-07","biren-shanghai-lab-2026-09-10","enflame-ipo","hbm-reuters","duv-reuters","mlperf-training","mlperf-inference","inferencex"]}]
Added: Biren's illustrated 2026 half-year results summary
Added: Biren reports GLM-5.3-Flash inference validation on Bili 166M
Added: Biren and Shanghai University establish a joint intelligent-chip laboratory
Added: CIOE field report on optical-module shipment ramps and supply constraints
Updated: Enflame’s Shanghai listing
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-12
Updated: Continuous inference benchmark dashboard
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-12
Updated: MLPerf Inference v6.0 submissions
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-12
Updated: MLPerf Training v6.0 submissions
1 field changes
- reviewed_at: 2026-09-11 → 2026-09-12

2026-09-11a519664 ↗

Align Observatory with Lab Index Astro, YAML, skills and owner Cloudflare Pages

Added: data/benchmarks.yaml
Added: A broader domestic ecosystem, with concentrated dependencies
Added: Qualification becomes the decisive test
Added: Capital expands; memory remains the constraint
Added: ACM Research Shanghai
Added: Shanghai Aishengna
Added: AMEC
Added: Biren
Added: Cambricon
Added: CIOMP
Added: CXMT
Added: Empyrean
Added: Enflame
Added: Eoptolink
Added: Beijing Guowang
Added: H3C
Added: Huawei
Added: Hwatsing
Added: Iluvatar CoreX
Added: InnoLight
Added: JCET
Added: Kingsemi
Added: Kunlunxin
Added: MetaX
Added: Moore Threads
Added: NAURA
Added: RSLaser
Added: SIOM
Added: Skyverse
Added: SMEE
Added: SMIC
Added: Alibaba T-Head
Added: YMTC
Added: Yuliangsheng
Added: Fujian Zhiqi Photonics
Added: data/layers.yaml
Added: HBM milestone reported at CXMT
Added: Initial domestic immersion DUV tools enter production
Added: Domestic lithography effort extends across components
Added: Enflame enters Shanghai’s public market
Added: EUV prototype reported, with chip production still ahead
Added: Field study documents Ascend integration work
Added: Huawei publishes FP4 pretraining research
Added: Memory constraints reach accelerator pricing
Added: Alibaba announces Zhenwu M890
Added: Optical modules retain international component links
Added: Teardown identifies Enflame S60 manufacturing route
Added: Teardown characterizes SMIC N+3
Added: MI355X
Added: TPU v7 / Ironwood
Added: Ascend 910C
Added: Atlas 350 / 950PR
Added: Blackwell Ultra / GB300
Added: data/overview.yaml
Added: data/research-state.yaml
Added: Semiconductor equipment portfolio
Added: Company and equipment portfolio
Added: Independent Ascend deployment field study
Added: Ascend hardware and software portfolio
Added: Atlas 350 accelerator specifications
Added: GPU company portfolio
Added: MLU and NeuWare portfolio
Added: Institute research
Added: Serving LLMs on Huawei CloudMatrix384
Added: CloudMatrix384 and GB200 system comparison
Added: CXMT’s reported HBM3E progress
Added: DRAM portfolio
Added: Huawei and the domestic lithography supply chain
Added: Initial domestic immersion DUV production
Added: EDA portfolio
Added: Enflame’s Shanghai listing
Added: Enflame S60 compute-chiplet analysis
Added: 1.6T transceivers using Broadcom DSPs
Added: Optical networking portfolio
Added: China’s EUV prototype program
Added: GB300 NVL72 specifications
Added: Optical-system patent
Added: Data-center networking portfolio
Added: HBM shortages raise Chinese accelerator prices
Added: HiFloat4 language-model pretraining study
Added: CMP and related process equipment
Added: Compute platform portfolio
Added: Continuous inference benchmark dashboard
Added: Optical networking portfolio
Added: XDFOI packaging manufacturability research
Added: Coating and developing equipment
Added: AI accelerator portfolio
Added: Zhenwu M890 announcement
Added: MI355X specifications
Added: MLPerf Inference v6.0 submissions
Added: MLPerf Training v6.0 submissions
Added: GPU and MUSA portfolio
Added: Process equipment portfolio
Added: Laser and optics research
Added: Inspection and metrology portfolio
Added: Lithography equipment portfolio
Added: SMIC N+3 process analysis
Added: T-Head chip portfolio
Added: TPU7x Ironwood specifications
Added: Foundry process portfolio
Added: NAND and enterprise storage portfolio