If it were up to the U.S. government, Beijing would not have access to advanced AI chips. Yet Chinese institutes own Nvidia GPUs, which were included with servers from Supermicro, Dell and Gigabyte.
Reuters found 10 Chinese universities and research institutes that obtained the AI hardware. The tenders took place between Nov. 20 and Feb. 28 but had not been previously reported. Unfortunately, Reuters does not report the exact make of GPUs involved. With values per tender between converted 10,000 euros and 250,000 euros, it could possibly involve different models.
Very limited quantities
The purchases are said to have involved very limited quantities of GPUs. Moreover, according to Nvidia, the chips were available well before the restrictions went into effect. Taking that at face value, it is likely that the Chinese institutions obtained at most a few last-gen A100 GPUs. Its successor, the H100, has always been scarce. Currently, America limits the maximum computing capacity to such an extent that even Nvidia’s strongest consumer GPU, the RTX 4090, may not be sold in China.
All the resellers mentioned deny that any rules have been violated. It is possible that the servers sold were still from old stock. In that case, the hardware should have been shipped before November, when the restrictions took effect.
Specific purposes for the chips are easy to guess. The Chinese customers are mostly research centers and focus on aerospace, among other things. Since GPUs are above all suitable for highly parallelized workloads, such scientific applications are extremely suitable. After all, they often involve countless calculations that are each relatively easy to process, making the thousands of cores in a GPU more suitable than a CPU with at most a few dozen cores.
Also read: Advanced Huawei chip doesn’t impress US Commerce Secretary