China Develops Its Most Powerful General-Purpose GPU: Meet Biren BR100, Faster Than NVIDIA Ampere In AI Horsepower With 77 Billion Transistors on 7nm
Birentech, a small enterprise based in Shanghai, China, has released the country’s most powerful General-Purpose GPU, the Biren BR100.
China Makes Its Most Powerful General-Purpose GPU To Date, The Birentech BR100 With 77 Billion Transistors
The Birentech BR100, which employs a proprietary GPU architecture and a 7nm production node, is the top general-purpose GPU on the market in China. Its die has 77 billion transistors. The GPU was created using TSMC’s 2.5D CoWoS architecture and has PCIe Gen 5.0 compatibility, 64 GB of HBM2e with a memory bandwidth of 2.3 TB/s, and 300 MB of on-chip cache (CXL interconnect protocol).
Brientech revealed a number of chip performance numbers during the launch. Based on the performance numbers, it appears that this chip will be quicker than the NVIDIA Ampere A100, at least on paper. It delivers up to 2048 TOPs (INT8), 1024 TFLOPs (BF16), 512 TFLOPs (TF32+), and 256 TFLOPs (FP32). In terms of the same GPU performance measures, the Hopper H100 GPU provides almost a 2x or 2.5x increase in performance. Additionally, the chip supports 512- and 64-channel encoding.
It’s interesting to note that the BR100 and NVIDIA H100 are comparably close in terms of overall transistor count. While the BR100 is only 3 billion transistors behind the 7nm production node, the H100 has 80 billion transistors on the new N4 process node. Thus, the die size would increase significantly.
The China-based business has introduced several chips in addition to the Biren BR100. Additionally, the Biren104 delivers half the performance metrics of the BR100, but the details of its specs have not yet been disclosed. The only information that is known about the second chip is that, unlike the Biren BR100, which employs a chiplet architecture, the BR104 is a monolithic die and has a TDP of 300W. It also comes in a typical PCIe form factor.
The manufacturer claims that a chip with 77 billion transistors may simulate the nerve cells in the human brain. Since the device will be utilized for DNN and AI applications, it will largely replace China’s reliance on NVIDIA’s AI GPUs.