Bloomberg reports that Taiwan Semiconductor Manufacturing Co. has suspended shipments of products it makes for China-based Biren, which designs processors aimed at artificial intelligence (AI) and high-performance computing (HPC) applications. TSMC's lawyers reportedly told the top contract chipmaker to suspend shipments while they are assessing the new limitations.
Keep in mind also that this information comes from a reputable yet unofficial source. So take it with a grain of salt.
The U.S. has imposed pretty strict curbs and sanctions of China's supercomputer and AI industries. Specifically, new chips cannot enable machines with the performance of over 100 FP64 PetaFLOPS, or over 200 FP32 PetaFLOPS within 41,600 cubic feet (1178 cubic meters). Furthermore, the machine cannot have a throughput of more than 600 GB/s.
Row 0 - Cell 0 | Biren BR104 | Biren BR100 | Nvidia A100 | Nvidia H100 |
Form-Factor | FHFL Card | OAM Module | SXM4 | SXM5 |
Transistor Count | ? | 77 billion | 54.2 billion | 80 billion |
Node | N7 | N7 | N7 | 4N |
Power | 300W | 550W | 400W | 700W |
FP32 TFLOPS | 128 | 256 | 19.5 | 60 |
TF32+ TFLOPS | 256 | 512 | ? | ? |
TF32 TFLOPS | ? | ? | 156/312* | 500/1000* |
FP16 TFLOPS | ? | ? | 78 | 120 |
FP16 TFLOPS Tensor | ? | ? | 312/624* | 1000/2000* |
BF16 TFLOPS | 512 | 1024 | 39 | 120 |
BF16 TFLOPS Tensor | ? | ? | 312/624* | 1000/2000* |
INT8 | 1024 | 2048 | ? | ? |
INT8 TFLOPS Tensor | ? | ? | 624/1248* | 2000/4000* |
Formally, Biren Technology (a fabless company with some Nvidia engineers behind it, which was valued at around $2.7 billion just last August) does not disclose all performance numbers of its compute processors. But the numbers it does disclose puts it on the map against Nvidia, Intel, and AMD. Just a reminder, Biren's GPUs cannot render graphics, so they aren't aiming to compete against the best graphics cards aimed at gaming.
Based on performance numbers published by Biren, its compute GPUs can barely compete with Nvidia's in HPC applications that require 64-bit precision for floating-point operations per second (FLOPS). What is more important is that Biren's software for AI and HPC is reportedly years and generations behind that of Nvidia. So whether or not Biren is a competitor to Nvidia as of today, restrictions or not, is still unclear.