AI GPU hardware and systems
As of 2026-09-11Biren · GPU company portfolio (source)Company dossier · Reviewed 2026-11-18
Biren 壁仞科技
General-purpose GPU compute contender.
What the evidence says
No verified latest-versus-latest performance ratio in the baseline research.
Biren · Biren's illustrated 2026 half-year results summary (source)Biren · Biren announces llm-d adaptation and reinforcement-learning framework integration (source)Biren · Biren reports GLM-5.3-Flash inference validation on Bili 166M (source)Shanghai Biren Technology · Placing of new H shares under general mandate (source)Shanghai Biren Technology / HKEX · Completion of placing of new H shares and amendments to articles (source)Biren · Biren and Shanghai University establish a joint intelligent-chip laboratory (source)Biren · Biren announces TorchSUPA PyTorch backend at ecosystem meetup (source)Biren · GPU company portfolio (source)llm-d · llm-d prefill/decode guide — Biren community compatibility configuration (source)RLinf and Biren · Embodied GRPO with RLinf and Biren GPUs (source)RLinf · Biren SUPA accelerator support merged into RLinf (source)Critical facts
RLinf PR 1581 merged on September 20, adding device discovery, visibility, BCCL communication and runtime installation. Its author reports a PI0.5 SFT end-to-end test; this review did not reproduce it or establish performance parity.
As of 2026-11-18RLinf and Biren · Embodied GRPO with RLinf and Biren GPUs (source)RLinf · Biren SUPA accelerator support merged into RLinf (source)The llm-d guide documents Qwen2.5-72B-Instruct-GPTQ-Int8 on eight Biren 166M GPUs, split into one prefill and one decode replica with TP=4 each. KV transfer uses NixlConnector over RDMA with CPU buffering. The project labels this community-maintained, outside nightly end-to-end testing; it is configuration evidence, not independently measured performance or production reliability.
As of 2026-11-17llm-d · llm-d prefill/decode guide — Biren community compatibility configuration (source)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.
As of 2026-06-30Biren · Biren's illustrated 2026 half-year results summary (source)Reports functional and accuracy validation of GLM-5.3-Flash on Bili 166M using SGLang and BIRENSUPA; no matched hardware speedup established.
As of 2026-09-07Biren · Biren reports GLM-5.3-Flash inference validation on Bili 166M (source)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.
As of 2026-09-07Biren · Biren announces TorchSUPA PyTorch backend at ecosystem meetup (source)September 2 Shanghai University laboratory agreement includes planned AI-chip optical-interconnect and cooling research, not production qualification.
As of 2026-09-02Biren · Biren and Shanghai University establish a joint intelligent-chip laboratory (source)Company says an unnamed new-generation product completed bring-up and entered post-silicon validation; customer sampling and volume production remain expected over coming quarters, subject to satisfactory validation and qualification.
As of 2026-10-08Shanghai Biren Technology · Placing of new H shares under general mandate (source)Proposed best-efforts placement of 130 million H shares at HK$31.08 each could raise an estimated HK$4.0197 billion net; 70% is allocated to supply-chain procurement, production readiness and commercialization, but the transaction had not completed.
As of 2026-10-08Shanghai Biren Technology · Placing of new H shares under general mandate (source)The October 12 filing confirms completion of 130 million H shares at HK$31.08, with HK$4.0404 billion gross and approximately HK$4.0197 billion net proceeds. This closes the previously conditional October transaction; funding availability does not establish completed procurement or qualified chip production.
As of 2026-10-12Shanghai Biren Technology / HKEX · Completion of placing of new H shares and amendments to articles (source)Company news
Timeline →RLinf and Biren document SUPA integration and embodied RL work
The co-authored account describes LingBot-VLA GRPO adaptation and numerical alignment. The linked SUPA accelerator-manager PR was merged on September 20; the model integration was still under review at publication.
Why it matters. Public upstream support is verifiable integration evidence. Co-authored optimization claims are not an independent cross-hardware benchmark; BRTT open-sourcing remained planned. This follow-up preserves the earlier October vendor announcement.
RLinf and Biren · Embodied GRPO with RLinf and Biren GPUs (source)RLinf · Biren SUPA accelerator support merged into RLinf (source)Biren completes October placement with HK$4.0197 billion net proceeds
Biren filed completion on October 12 of the previously announced 130-million-share placement at HK$31.08 per share, raising HK$4.0404 billion gross and approximately HK$4.0197 billion net.
Why it matters. Financing completion advances the earlier supply-chain and production-readiness funding plan. It is the same October transaction, not an additional raise; the October 8 conditional record and product-validation expectations are preserved.
Shanghai Biren Technology / HKEX · Completion of placing of new H shares and amendments to articles (source)Biren describes distributed inference adaptation and RL integrations
Biren says it submitted a 166M adaptation to llm-d with Dynamia and integrated RLinf, slime and verl support into community mainlines. Its inference example separates prefill and decode across two four-GPU nodes using RDMA and CPU-buffered KV transfer.
Why it matters. This backfilled announcement is vendor integration evidence, not a matched performance comparison. The llm-d guide checked on November 17 documents a community compatibility configuration outside nightly end-to-end testing; that later check does not establish the guide's October contents. The three RL mainline claims were not independently audited here.
Biren · Biren announces llm-d adaptation and reinforcement-learning framework integration (source)llm-d · llm-d prefill/decode guide — Biren community compatibility configuration (source)Biren says its new-generation product completed bring-up
Biren's placement filing says its unnamed new-generation product completed bring-up and entered post-silicon validation. It expects customer sampling and volume production over the coming quarters, subject to satisfactory validation and qualification.
Why it matters. This is an engineering milestone, not customer sampling, qualification or volume production. The filing does not name the product or disclose specifications, fabrication, memory, yield, customers, workload benchmarks, production volume or a firm sampling schedule.
Shanghai Biren Technology · Placing of new H shares under general mandate (source)Biren proposes a placement with estimated net proceeds of HK$4.02 billion
Biren agreed a best-efforts placement of 130 million new H shares at HK$31.08 each, for gross proceeds of HK$4.0404 billion and estimated net proceeds of HK$4.0197 billion. It allocates 70% of net proceeds to strategic supply-chain procurement, production readiness and commercialization, 20% to research, software and ecosystem development, and 10% to working capital.
Why it matters. This is a conditional financing plan, not completed financing or production. Listing approval remained a prerequisite and the filing warned that the placement might not proceed. As of September 30, Biren had used HK$982.8 million, or 14%, of its prior July placement proceeds; the filing does not disclose suppliers, reserved capacity, counterparties or commercial commitments.
Shanghai Biren Technology · Placing of new H shares under general mandate (source)Biren and Shanghai University establish a joint intelligent-chip laboratory
Biren reports a September 2 agreement with Shanghai University. Planned joint research includes AI-chip optical interconnects and cooling technologies.
Why it matters. Adds a research and talent-development channel; the agreement does not demonstrate a qualified optical interconnect or production cooling system.
Biren · Biren and Shanghai University establish a joint intelligent-chip laboratory (source)Biren's half-year summary reports revenue growth alongside continuing losses
Biren's published infographic reports first-half 2026 revenue of RMB 1.236 billion, up 1,997.6% year on year, and a loss for the period of RMB 377 million. R&D expenditure was RMB 804 million. It also reports supplier qualification and batch deliveries to major internet customers without naming them in the infographic.
Why it matters. More concrete commercialization evidence than a product announcement, but company-reported revenue does not establish customer utilization, domestic manufacturing provenance or benchmark parity.
Biren · Biren's illustrated 2026 half-year results summary (source)Biren reports GLM-5.3-Flash inference adaptation on Bili 166M
Biren says SGLang and BIRENSUPA on Bili 166M passed functional and accuracy checks for GLM-5.3-Flash, including loading, serving, text question-answering and image understanding.
Why it matters. A concrete software-compatibility milestone, not an independent throughput, latency, power or training time-to-quality comparison. The announcement's Day0 branding does not establish the exact validation date.
Biren · Biren reports GLM-5.3-Flash inference validation on Bili 166M (source)Biren announces TorchSUPA PyTorch backend plugin
Biren's September 7 account of an August 28 ecosystem meetup says it released TorchSUPA, a PyTorch PrivateUse1 backend plugin for Biren GPUs, with support for common APIs, collective communication and FSDP2, graph capture and profiling.
Why it matters. A public backend can reduce framework-integration and migration friction, but Biren's claims about CUDA-style code reuse do not establish operator coverage, correctness, performance, upstream acceptance or CUDA parity. No public code repository was identified in this review.
Biren · Biren announces TorchSUPA PyTorch backend at ecosystem meetup (source)