
American authorities last year introduced a ban on the export of graphics processors to China with the level
What is known
From the names you can understand that the A800 isa slower version of the A100, while the H800 is a slower version of the H100. Demand for such chips has grown due to a surge in interest in generative models like the Generative Pre-trained Transformer (GPT).
As it became known, the performance of the A800 compared to the A100 fell by 30%. You can see the results below:
- FP64 - 6.8 TFLOPS vs 9.7 TFLOPS;
- FP64 using Tensor Cores – 13.7 TFLOPS vs. 19.5 TFLOPS;
- FP16 and BF16 - 439 TFLOPS vs 624 TFLOPS.
The A800 costs almost $14,500 in China. As for the higher-performance H800 chip, there is no information yet on how much NVIDIA has reduced its performance compared to the H100.