Nvidia has developed a new data center GPU for the Chinese market that complies with updated US export control regulations. The chip, known as the A800, can be utilized in place of the A100 for identical functions, according to a report from Reuters. The A100 is a Tensor core GPU based on the Ampere architecture used in supercomputers by companies like Oracle, Tesla, and Meta(as in the RTX 30 series). They are extensively utilized for high-performance computing.
They offer up to 80GB of memory per chip, the fastest bandwidth in the world at 2TB/s, and PCIe Gen 4 connectivity, allowing thousands of GPUs to operate simultaneously. Advanced and innovative exports to China are prohibited by US trade regulations put in place earlier this year. To avoid such cutting-edge items ending up in the hands of the military and being utilized for gathering intelligence through satellite photos or digital communications. The guidelines expressly cap the chip transfer rate at less than 600GB/s.
A 400GB/s transfer rate is available with the new A800 GPU. The rules don’t just apply to Nvidia; AMD told Reuters in September that it could no longer sell its MI250 chips, which are also used in supercomputers, to the large Asian nation but thought its MI100 one HPC GPUs were unaffected. If an exception was not granted, it was alleged in the same source that Nvidia would lose $400 million in just one quarter. The new chip is designed to get around this.
China claims that the US is imposing a “tech blockade” that threatens the continuity of international supply chains. In September, the spokesperson for the Chinese commerce ministry, Shu Jieting, stated that “the US continues to abuse export control procedures to restrict exports of semiconductor-related commodities to China.” The A800 can’t be programmed to go beyond the clear standard for decreased export control set by the US government.