China with extremely powerful GPU in 7nm process
Birentech has introduced an all-purpose GPU that is said to beat Nvidia’s A100 in terms of its raw computing power and is manufactured in a 7nm process. The graphics card is equipped with 64GB of HBM2e memory and is only available in the OAM form factor. Only the H100 from Nvidia should clearly outperform the Chinese graphics card. Read more about this below.
Special production and development of semiconductor elements provides a certain amount of safety, especially in times of shortages and crises, and now that Russia has begun to create “its own” chips and laptops (we mentioned), it is no wonder that China has already taken several steps is further It has provided a very powerful GPU for AI computations that should clearly outperform Nvidia’s Ampere-based A100. The GPU is manufactured by Birentech, a small company from Shanghai.
Five years of Radeon Vega: the last generation of GCN with up to 16GB of HBM [Update: RX Vega wird fünf]
The Berain BR100 itself has nothing to do with the compact size. 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 in a TSMC 1.5D CoWoS design and includes 300 MB cache and 64 GB HBM2e memory with a memory bandwidth of 2.3 TB/s and native support for PCI-E 5.0.
In the presentation, the manufacturer addressed the performance of the new flagship. The BR100 should achieve 2048 TOPs (INT8), 1,024 TFLOPs (BF16), 512 TFLOPs (TF32+), and 256 TFLOPs (FP32). According to 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). It is not possible to verify the validity of these metrics yet, but if they are correct, then Nvidia’s A100 is beaten and only the H100 offers 2 to 2.5 times the performance of a Chinese GPU.
Compared to the H100, the small difference in transistors should mean that the GPU is correspondingly bulky and expensive. H100 actually measures 814 mm², but it is made in TSMC’s 4nm process and not at 7nm. Birentech limits the TDP to 550W and the graphics card will only be offered as an OAM module, while it will soon follow a reduced expansion phase as a PCI-E model with a TDP of 300W.
Lifelong foodaholic. Professional twitter expert. Organizer. Award-winning internet geek. Coffee advocate.