Chinese Chip Company Releases 77 Billion Transistor BR100 GPU To Rival NVIDIA Hopper H100
China's very eager to not only rid itself of dependence on foreign silicon but also to display dominance of its domestic processors. That's likely the impetus behind some of the more bombastic claims we've seen from Chinese chipmakers, like the Fantasy One
GPU. Innosilicon recently debuted
its Fantasy Two GPU that seems much more down-to-earth, but it's hardly the only company making such claims.
Have you ever heard of Biren Technology? Established in 2019, the company was created to focus on design and creation of domestic compute accelerators for the Chinese market. Even though we say "GPU", it's fairly unlikely that Biren's new chip actually has any graphics capabilities at all; it appears to be completely focused on GPGPU-style compute.
We were initially skeptical upon reading of Biren and the BR100. After all, who wouldn't be skeptical when hearing about a brand-new company debuting a processor that can compete with NVIDIA's A100
in AI workloads? Naturally, there are some caveats to this news, not least of which is that, as far as we can tell, Biren has not actually publicly demonstrated the chip yet. As such, this is basically a paper launch of the 77b-transistor BR100 processor.
Biren Technology claims that the BR100 can reach 2048 TOPS—that's 2 exa-ops—in INT8 compute, with performance purportedly scaling down in linear fashion by data size. That results in a claim of 256 TFLOPS FP32 compute, something which makes even NVIDIA's H100
(and its 60 TFLOPs spec) look puny.
BR100 apparently achieves this performance through the use of 64GB of HBM2E memory, 300MB of on-die cache, and a chiplet-based design built using 2.5 CoWoS fabrication on TSMC's 7nm process. The BR100 itself comes on a mezzanine card with a beefy heatsink, not unlike NVIDIA's fastest compute accelerators. Its Open Accelerator Module design uses Compute Express Link
(CXL) to connect to the host system at 2.3 TB/sec.
Biren Technology also debuted the BR104 processor, which we're guessing is probably a single chiplet from the dual-chiplet BR100 design. It comes in standard PCIe card form factor, and its creators claim it can provide exactly half the performance of the BR100 within a 300W add-in card form factor. Biren says the chips are designed for medical imaging, molecular dynamics, and other physical simulations.
Biren did not come out of nowhere overnight; the company has been collecting funding since its founding in 2019 and announced a successful test
of the BR100 in April of this year. While the company doesn't seem to have done a live demo of BR100 yet, it seems legitimate enough to us. The question is whether the BR100 and its BiLiren architecture can live up to its creators' ambitious claims.
We will probably find out more about these processors when the company presents at Hot Chips
34, starting August 21st.