Performance evidence · Reviewed 2026-11-02

What has actually been measured?

No verified matched frontier end-to-end comparison

Follow-up through November 2 did not verify a matched independent end-to-end comparison of current Chinese accelerators against Rubin, GB300, MI355X, TPU v7, or OpenAI Jalapeño. The review checked MLPerf Training v6.0 and Inference v6.1 repository heads (unchanged since the previous submitter review) plus InferenceX's published coverage, not every result artifact. InferenceX independently analyzed a 16-rank Ascend 950DT DeepSeek V4 profiler trace and confirmed functioning optimized day-one inference, but did not publish its planned same-harness H200/B200 numerical comparison. MLPerf's standardized results are submitter-run and are not automatically independent measurements. MLPerf Inference v6.1 advanced to commit 4bb63cd28eb136be221bc801a9e38c66c8a7553f with previously missing submission documentation; its 30 closed submitter directories still contain no current Chinese accelerator vendor. 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.

Published hardware specifications

Dense BF16 peak per marketed accelerator and HBM bandwidth. These are specifications, not model training or inference benchmarks. Package design and power differ.

Scroll sideways to compare all columns →

AcceleratorDense BF16
(PFLOP/s)
HBM bandwidth
(TB/s)
Basis & limitations
MI355XAMD · Outside China2.58Dense BF16 peak; high-power accelerator module.AMD · MI355X specifications (source)
TPU v7 / IronwoodGoogle · Outside China2.3077.38Per marketed chip, containing two compute chiplets.Google Cloud · TPU7x Ironwood specifications (source)
Ascend 910CHuawei · China0.7523.2Dual-die accelerator package; specifications from the 2025 paper.Huawei / SiliconFlow · Serving LLMs on Huawei CloudMatrix384 (source)
Atlas 350 / 950PRHuawei · China0.4251.4Inference/prefill-oriented PCIe card, up to 600 W.Huawei · Atlas 350 accelerator specifications (source)
Blackwell Ultra / GB300NVIDIA · Outside China2.58Per GPU, derived from 72-GPU system totals; sparsity removed from BF16 peak.NVIDIA · GB300 NVL72 specifications (source)

DeepSeek-R1 on Ascend 910C and H100

2025-06 · vendor study · Historical comparison

Scroll sideways to compare all columns →

MetricAscend 910CH100 baseline
Prefill tokens/s per accelerator5,6556,288
Decode tokens/s per accelerator1,9432,172
Decode time per output token49.4 ms55.6 ms

Vendor/partner measurements versus published NVIDIA baselines. INT8 versus FP8, different decode batches, and speculative-decoding assumptions prevent a universal ratio. Perfect expert-balancing results are projections.

Huawei / SiliconFlow · Serving LLMs on Huawei CloudMatrix384 (source)

Benchmark evidence register

2026-10-26vendor

MetaX RLinf embodied training

MetaX reports 1,024-GPU Pi0.5/RoboTwin SFT: 20,000 steps in 32 hours, 2.75 samples/s per GPU versus 2.8 at 512 GPUs, using RoCE and GPFS. Its separate greater-than-97% scaling claim covers ManiSkill asynchronous PPO from 8 to 64 GPUs, not the thousand-GPU SFT run. A reported task-success improvement used a larger global batch and newer framework; it is not a hardware speedup. No matched foreign time-to-quality, energy or recovery comparison was supplied. Upstream PR checks do not independently reproduce these cluster results.

MetaX · MetaX announces RLinf upstream integration and thousand-GPU validation (source)RLinf contributors · RLinf PR 1656: MetaX MACA support and validation limitations (source)
Reviewed 2026-10-27; publication date unverifiedmeasured

ubenchX core memory-access profiles

SemiAnalysis publishes scalar pointer-chase memory profiles for Ascend 950PR_9579 (128 MiB allocation, Cube and Vector controllers) and B200. Values measure differences between matched-address profiles, not direct core-to-core latency or physical distance. Ascend SYS_CNT ticks and B200 clock64 cycles are different units and cannot support a cross-device latency ratio. Concurrent launches lack a grid-wide starting barrier; fine grouping is less repeatable and does not establish physical cache sharing. These are hardware microbenchmarks, not end-to-end AI performance; the Observatory did not rerun them.

SemiAnalysis / InferenceX · ubenchX core memory-access profiles: Ascend 950PR and B200 (source)
2026-10-20vendor

DeepInfer published performance scope

DeepSeek publishes prefill and decode throughput/latency plots using 4 and 64 NVIDIA SM100 GPUs respectively, each within one scale-up domain. Ascend 950 is a supported software target, but these plots are not Ascend measurements or a matched cross-hardware comparison. The offline release excludes complete service infrastructure. No benchmark was rerun by the Observatory.

DeepSeek-AI · DeepInfer initial release: supported platforms, service exclusions and benchmark hardware (source)
2026-10-21measured

Ascend 950PR kernel study

SemiAnalysis reports a best GEMM kernel reaching 99.7% of its compute ceiling; below the compute-bound ridge, median achieved/attainable performance was 41.5% versus H100's 46.8%. Power and clock limits differed. High kernel utilization is not model utilization. Its EP16 communication comparison with GB300 includes different operator work, precluding an unqualified speedup.

SemiAnalysis — Bryan Shan and Dylan Patel · The Huawei-DeepSeek Symbiosis: Ascend 950PR microbenchmarks (source)
2026-07-22measured

Ascend 910 scientific-computing study

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.

Pengcheng Laboratory · Ascend to Science: Exploration of AI Chips for Scientific Computing (source)
Continuousmeasured

InferenceX

Reproducible independent inference runs. A June 9 article examined a 16-rank Ascend 950DT DeepSeek V4 profiler trace and verified working day-one inference, but provided no public same-harness numerical comparison with H200 or B200; the promised follow-up was not found. The current dashboard names Rubin, GB300, MI355X, TPU v7 and OpenAI Jalapeño systems, without a verified matched current Chinese result.

SemiAnalysis InferenceX · DeepSeek V4 1.6T performance over time and Ascend 950DT analysis (source)SemiAnalysis · Continuous inference benchmark dashboard (source)

How to read these results

Independent means the evaluator ran the workload independently of the vendor. MLPerf provides standardized rules but its submissions are generally run by submitters. Vendor and partner studies remain attributed to their authors.

A useful comparison aligns model, quality, precision, context, batch, latency target, accelerator count, software and power. Training needs time-to-quality and reliability; inference needs both throughput and latency. Missing evidence does not establish incapability.