Birentech challenges NVIDIA with the most powerful Chinese GPU ever

A little over a year ago, we reported that Chinese manufacturer Birentech was preparing to launch its first GPU built at 7nm, which would compete for the compute GPU market alongside rivals AMD and NVIDIA. A few months later, the company officially presented Biren BR100, the most powerful general-purpose GPU ever made in China, which can count on 77 billion transistors, only 3 billion fewer than NVIDIA Hooper H100 (which however was built on the new process node N4).

As anticipated, the chip, built on a 7nm process node, uses TSMC's 2.5D CoWoS design and features 300MB of on-chip and 64GB cache of HBM2e with a memory bandwidth of 2.3TB / s, as well as support for PCIe 5.0 (CXL). As for performance, we are talking about a maximum of 2.048TOPS (INT 8), 1.024TFLOP (BF16), 512TFLOP (TF32 +) and 256TFLOP (FP32); respectable numbers, even higher than NVIDIA's Ampere A100, but still far from Hopper H100. The GPU will be available on boards with OAM form factors and will use the passive tower cooling solution made by Birentech itself.



Photo Credit: BirenTech
In addition to the BR100, Birentech will also offer the BR104 chip, which will only be half as powerful as its big brother, although full specifications have not been officially disclosed. Another substantial difference lies in the design, which in the case of BR104 will be the traditional monolithic die, while BR100 employs a more modern chiplet organization. Finally, BR104 will be available only with the standard PCIe form factor and the card will have a TDP of 300W.