According to Reuters, the cut-down NVIDIA A100, also known as the A800, serves as an “alternative product” to the original A100 GPU and meets the US government’s test for reduced export control. To that end, the GPU cannot be programmed to exceed the limitations placed on it. While not an ideal alternative, not doing anything to keep the supply chain moving would cost NVIDIA hundreds of millions of dollars in revenue. Specs-wise, the A800 has a chip-to-chip data transfer rate of 400GB/s, down from 600GB/s on the A100. For context, the new export restriction for supercomputer GPUs prohibits chips with 600GB/s and up from being shipped out to China or Russia.
As a quick recap: back in August, both NVIDIA and AMD had their advanced chips, including the A100, added to the export control list by the US Commerce Department. Months later in October, Washington instated a ban on the export of advanced microchips and equipment to Chinese chipmakers, in an effort kneecap the country’s semiconductor industry and in turn, the expansion and advancement of its military. On another note, and to be fair, the US government’s fears aren’t unfounded. Theoretically speaking, if China did have access to the full-fat A100 GPU, there is a high likelihood that – just as most countries with an advanced military infrastructure would – it would be used for more nefarious purposes. Among them are the fears that Beijing could use them to create “automated targeting systems for drones” or further its already advanced facial recognition infrastructure, used in the surveillance of its population. (Source: The Verge, Reuters)
