NVIDIA will not let the Chinese server and supercomputer market slip through its fingers. Two months after having seen two of its best GPUs being deprived of exports to China by the American authorities, the Californian giant is back with a chip designed precisely… to circumvent these sanctions.
” Another alternative product to NVIDIA A100 GPU for our customers in China », it is in these terms that NVIDIA soberly describes its A800 GPU, a less efficient substitute than the A100, specifically designed to circumvent American sanctions and be distributed without hindrance in China. ” The A800 meets clear U.S. government criteria for reduced export controls and cannot be programmed to exceed them “, specifies a spokesperson for the firm in a press release addressed to TechCrunch.
Chinese supercomputers have an NVIDIA solution cut out for them
The A800 therefore makes up for the impossibility of selling A100s in China. A vaguely roundabout way to still reach the Chinese market for supercomputers, artificial intelligence and data centers for which the A100 was developed.
As a reminder, the A100 and the H100 (designed more specifically for businesses) were placed in early September on the list of products whose exports to China are tightly controlled by the United States. A measure taken by the American authorities in an attempt to ” address the risk of these products being used, or diverted, to a military end-use or military end-user in China and Russia “.
Offering Chinese companies a replacement for these two chips is crucial for NVIDIA. As the firm had specified in September, the banishment of the A100 and H100 in China represents a shortfall of nearly 400 million dollars in the third quarter of 2022 alone. This explains the promptness with which NVIDIA has developed and launched, in this same third quarter, its A800.
A100, A800 what differences?
As specified TechCrunch, most of the main technical specifications of the A100 were kept to result in the A800, except for the interconnect speeds. We thus learn that the A800 operates at “only” 400 Gbps while the A100 operates at 600 Gbps per second. By restricting this characteristic, NVIDIA ensures that it stays just below the maximum threshold set by the American authorities… and thus to be able to market the A800 without any problems.
If this limitation applies to this particular specification, it is for a very specific reason described by the Center for Strategic and International Studies, a think tank bipartisan, precise TechCrunch.
” By targeting only chips with very high interconnect speeds, the White House is trying to limit its controls to chips designed to be networked in data centers or supercomputers that train and run large AI models. “, we read.
Source : TechCrunch
23