NeuroAI NEUROAINEUROAI.SITE
ESC

Biren's BR100: a Chinese GPGPU that beat the A100 — then hit the wall

The story of Biren's BR100 and BR104, the TSMC node they were built on, and how US sanctions forced a painful rebuild on home soil.

2026-10-03 · 885 words · NeuroAI
Biren's BR100: a Chinese GPGPU that beat the A100 — then hit the wall

A startup staffed by ex-Nvidia engineers unveils a chip that, on paper, outruns one of the West's best. Weeks later, the factory that was making it is told to stop. The most instructive part of the story is not the benchmark — it is what happened next.

The chip that turned heads

At its 2022 launch event, Shanghai-based Biren (壁仞科技) presented the BR100, a general-purpose GPU (GPGPU) built on TSMC's 7-nm process using advanced 2.5D CoWoS packaging. It packed roughly 77 billion transistors and was paired with 64 GB of HBM2e memory. Biren claimed peak throughput around 1,000 TFLOPS of BF16 and over 2,000 TOPS of INT8 — numbers that, on paper, edged Nvidia's then-current A100.

A companion part, the BR104, offered about half the compute, 32 GB of memory, and a 300 W power envelope in a standard card form. Together they were meant to seed a domestic alternative to Nvidia's data-center lineup at a moment when Chinese buyers were starting to worry about supply.

The wall it hit

The impressive specs came with a fragile foundation. The BR100 was manufactured at TSMC, and in October 2022 new US export controls on advanced computing effectively halted that production. A year later, in October 2023, the US Commerce Department added Biren and several of its subsidiaries to the Entity List, citing the development of advanced computing integrated circuits. The practical effect: Biren lost access to TSMC's leading nodes, to US-origin EDA tools, and to many foreign items built with American technology.

For a chip designer, that is not a speed bump. It is a rebuild.

The pivot to a domestic stack

Biren's response was to redesign its products for SMIC's (中芯国际) 7-nm-class nodes and to assemble what industry analysts now describe as a largely domestic supply chain — domestic foundry, domestic advanced packaging to replace TSMC's CoWoS, and domestic HBM sources whose names are not publicly disclosed. Reported follow-on parts such as the BR106 and BR166 are framed as optimized for that homegrown process. Biren has also been reported to be preparing a Hong Kong listing to fund the rebuild, though details remain tentative.

The lesson is sobering. A brilliantly designed chip is only as sovereign as the weakest link in its supply chain. Biren's BR100 proved Chinese teams could architect at the frontier; the sanctions proved that architecture alone is not enough.

The context around the chip

Biren was founded in 2019 by Lingjie Xu (徐凌杰) and a team drawn from Nvidia, Alibaba, and STMicroelectronics — exactly the kind of returnee talent China's chip programs are built around. The BR100 did not arrive in a vacuum; it won the highest honor ("镇馆之宝") at the 2022 World Artificial Intelligence Conference in Shanghai, a signal of how much national expectation rode on it. That makes the subsequent sanctions blow sharper: a chip celebrated as a symbol of self-reliance was, at launch, almost entirely dependent on a Taiwanese foundry and American tooling. The story of Biren is now less about a single product and more about whether China can indigenize the unglamorous middle of the stack — packaging, HBM, EDA — fast enough to matter.

What survives from the BR100 era

Two things from Biren's first act still matter:

  • A software platform, BIRENSUPA, that supports mainstream frameworks including PaddlePaddle and claims adaptation for over a thousand AI models — proof that ecosystem, not just silicon, is where Chinese GPU firms live or die.
  • A chiplet design philosophy: the BR100 was built from two compute dies, a choice that let Biren break reticle-size limits and reuse one tapeout for multiple products. That engineering approach travels even when the foundry does not.

What readers can do now

  • Judge domestic GPU firms by supply-chain sovereignty (foundry, packaging, HBM, EDA), not just launch-day TFLOPS.
  • Watch for Biren's BR106/BR166 量产 (mass-production) signals and any confirmed SMIC node — those confirm the rebuild is real.
  • Treat any "beats Nvidia" claim with the date in mind; the BR100's A100 comparison is from 2022 and predates both newer Nvidia parts and the sanctions pivot.

What Biren's story really shows

Biren is the clearest case study in how sanctions reshaped China's chip map. A design that looked competitive in 2022 became unbuildable overnight when its foundry access was cut, and the company's survival now depends on a domestic supply chain that did not exist when the BR100 launched. The lesson is not about one chip but about sovereignty: a leading-edge design is only as real as the foundry, packaging, HBM, and EDA that can actually produce it. For readers, Biren is worth following as a live test of whether China's semiconductor stack can close the last miles — and as a reminder that a spec sheet from 2022 says little about what ships in 2026.

Honest limitations

BR100 and BR104 specifications are drawn from Biren's 2022 launch materials and tech-media reporting (Notebookcheck, WCCFtech, EE Times), not from independent teardowns we performed. The BR100's performance claims were against Nvidia's A100, not later H20/H100 parts. The reported Entity List addition (October 2023) and TSMC production halt (October 2022) are documented by US regulatory filings and reputable media. Biren's current foundry, node, HBM supplier, and Hong Kong listing status are not officially confirmed in detail; we treat them as reported, not verified. BIRENSUPA's "1,000-model" claim is vendor-stated.

Related coverage

More in “Chips & Compute” → · Back to home · Markdown version