Nvidia Forced To Make A Special GPU For China Due To US New Rules

Nvidia created a special GPU for the Chinese market. (photo: dock. Nvidia)

JAKARTA - The new US export control rules force Nvidia to create a special GPU for the Chinese market. The chip, the A800, is claimed to be a replacement for the A100 but has the same task.

First reported by Reuters, US trade rules set earlier this year limited export activities of advanced chips or AI training equipment to countries such as China and Russia.

The rule was made in preventing the high-tech product from falling into the hands of the military and used for intelligence collection from digital communications or satellite imagery. It limits the speed of chip transfer below 600GB/dtk.

Meanwhile, the new A800 GPU has a transfer speed of 400GB/. The A800 GPU should be just as good for simple tasks but will be less effective when used on supercomputers.

An unnamed Nvidia spokesman said the GPU was an alternative product of the A100 GPU for customers in China that met clear tests from the US Government, in order to reduce export control and could not be programmed to exceed it.

Reporting from Tom's Hardware, Wednesday, November 9, the A100 GPU is widely used on supercomputers by Oracle, Tesla and Meta. This is a Tensor core GPU based on Ampere architecture.

GPU A100 is widely used for AI, and in-depth learning tasks such as image and speech recognition, as well as high performance computing. The GPU accommodates memories of up to 80GB per chip, the world's fastest memory bandwidth at 2TB/dtk, and the ability to connect together via PCIe Gen 4 which allows thousands of GPUs to work simultaneously.

Nvidia isn't the only company affected by the rules, AMD last September told Reuters it could no longer sell its MI250.

The chip is also used in supercomputers in densely populated Asian countries. In the same report, it is claimed Nvidia will lose US$400 million to Rp6.2 trillion in just a quarter if it is not granted an exception, something that seems to be made to be passed by the new chip.