Newly imposed US GPU export restrictions curtail AMD and other US semiconductor vendors serving China.

The U.S. Department of Commerce has notified AMD that it must obtain an export license to ship its Instinct MI250 and MI250X compute GPUs to China, marking a new restriction on high-performance computing equipment. AMD confirmed receipt of this alert in late August, the same timeframe Nvidia received similar documentation. Unlike Nvidia, which faces severe commercial impact, AMD's data center business will only be marginally affected since the company sells relatively few compute GPUs to China, with most shipments directed to supercomputers in the U.S. and Europe.

Nvidia stands to shoulder the heavier burden, potentially losing $400 million this quarter due to the restrictions on its A100, A100X, H100, and more advanced compute GPUs. The company has begun redirecting some orders toward its A30 compute GPUs as a workaround. The U.S. Department of Commerce imposed these restrictions because it does not want high-performance compute GPUs to fall into the hands of the Chinese military or associated government agencies, which could leverage supercomputers based on these chips to develop new weapons or optimize chip designs for military applications. Supercomputers used for weapons design now rely on both AI for pathfinding and HPC for simulations.

AMD's Instinct MI210, which offers considerably higher FP64 performance than Nvidia's A100 and can challenge the H100 in FP64 matrix operations, can still be sold to China without an export license, according to reports. However, the MI210 falls considerably behind in AI performance. Based on official performance numbers, the Department of Commerce appears more concerned with AI performance than HPC performance in determining which products require licensing.

Chinese companies including Alibaba, Baidu, and Tencent depend on high-end Nvidia chips for their AI services. Without access to these processors, they will be forced to rely on lower-performance A30 compute GPUs or turn to AI cloud instances from AWS or Google. A former AMD China executive told Reuters, "It is a resource impact. They will still work on the same projects, they will still be moving forward; it just slows them down." Meanwhile, Chinese firms are developing domestic alternatives—compute GPUs and AI accelerators such as Biren's BR100 and Baidu's Kunlun II can challenge Nvidia's A100 in terms of performance and H100 in terms of complexity.

While export license requirements for GPUs comparable to or exceeding Nvidia's A100 may slow certain projects in China, assuming the Department of Commerce does not approve some customers, the restrictions will not halt them completely. The combination of export controls and advancing domestic semiconductor capabilities suggests the Chinese technology sector will adapt, though potentially at a reduced pace.