China with extremely powerful GPU in 7nm process

Own production and development of semiconductor elements provides some security, especially in times of shortages and crises, and now that Russia has begun to build “own” chips and laptops (we reported), it is not surprising that China has already given several Steps. goes further and has presented a very powerful GPU for AI calculations that should clearly outperform Nvidia’s Ampere-based A100. The GPU was made by Birentech, a small Shanghai company.

Five years of Radeon Vega: the latest generation of GCN with up to 16 GiByte HBM [Update: RX Vega wird fünf]

The Biren BR100 itself has nothing to do with small. The GPU is manufactured using the current 7nm process and contains 77 billion transistors, just three billion fewer than Nvidia’s hopper-based H100. It is built on the TSMC 1.5D CoWoS design and includes 300 MiB cache, 64 GiB HBM2e memory with 2.3 Tb/s memory bandwidth, and native support for PCI-E 5.0.

Birentech BR100

Font: Wccftech



In the presentation, the manufacturer addressed the benefits of the new flagship. The BR100 should achieve 2048 TOP (INT8), 1024 TFLOP (BF16), 512 TFLOP (TF32+) and 256 TFLOP (FP32). Depending on the manufacturer, Nvidia’s A100 achieves 624 or 1,248 TOPS (INT8), 312 or 624 TFLOPs (BF16), 156 or 312 TFLOPs (TF32) and 19.5 TFLOPs (FP32). To what extent these metrics are true cannot be verified yet, but if they are correct, Nvidia’s A100 is outclassed and only the H100 offers 2 to 2.5 times the performance of the Chinese GPU.

Compared to the H100, the small difference in transistors must mean that the GPU is huge and consequently expensive to manufacture. The H100 already measures 814mm², but it’s manufactured on TSMC’s 4nm process and not 7nm. Birentech specifies the TDP at 550 Watts and the graphics card will only be offered as an OAM module, while a reduced expansion stage as a PCI-E model with a TDP of 300 Watts will soon follow.

Font: Wccftech