NVIDIA has released a cut-down version of its high-end A100 graphics chip for data centers to bypass restrictions recently imposed by the US government. They are needed to make it harder to sell machine learning equipment to countries like China and Russia. A cut-down version of the A100 is called the A800 and will have the same computing capabilities as the original, but will be able to exchange data with other cards in the computing cluster at a speed of 400 gigabytes per second instead of 600 GB/s. As reports The Wall Street Journal, this means that the A800 will be just as good at simple tasks, but will be significantly less efficient when part of something like a supercomputer.
An unnamed NVIDIA representative reported to Reuters that the card is an “alternative product to the NVIDIA A100 GPU for customers in China” that “meets the US government’s strict export control test and cannot be programmed to exceed it.” While the exact specifications of the US tests have not been disclosed, the WSJ says they cover interconnection speeds as well as “other performance thresholds.”
NVIDIA’s claims that GPUs cannot be programmed to exceed the limits are questionable. Back when the chips were in short supply, the company announced that its new cards would have limiters that could not be cracked, which would supposedly make them less desirable for cryptocurrency mining. As you know, soon users were able to bypass these restrictions.