NVIDIA already has new data-center GPUs for China NVIDIA responded quickly to new restrictions on exporting powerful GPUs from the U.S. Recently, NVIDIA ceased the shipment of high-performance GPUs to countries like China. The new U.S. restrictions aim to curb the utilization of powerful accelerators for specific tasks, particularly in machine learning and its potential […]
What exactly did US export limit? FP64 performance?
Fp8 performance per unit die apparently.
The 4090 has 660 fp8 tflops (which is insane when you think about it) and got the ban.
H100 is 1930 Fp8 tflops.
With sparsity both can do upto 2x that
The 4090 has more fp32 performance than the H100 though.
It’s not specifically fp8, but TOPS*data size. Absolute limit is 4800, or 5.8/mm^(2). Above either is an outright ban. Above half of either needs a license.
It’s slightly complex, as there are two metrics it needs to be under.
See this chart:
https://cdn.mos.cms.futurecdn.net/dHjnhPMk93HuDPBYnXBzLV.png
Total Processing Performance (TPP) is essentially the listed processing power multiplied by the length of operation (e.g., FLOPS or TOPS ‘8/16/32/64) without sparsity
Performance Density is counted by dividing TPP by the die area measured in square millimeters