In a recent report, the China Academy of Information and Communications Technology (CAICT) recommended that data centres in mainland China opt for Nvidia chips due to the high costs associated with transitioning to domestic solutions. The government-backed think tank suggested that data centres could consider Nvidia’s A100 and H100 high-performance computing units, or opt for H20 or other domestic alternatives for less demanding computing needs.
While Nvidia chips are popular in China, the company faces pressure from the US to restrict the sale of its GPUs to the mainland, citing concerns about potential military use. Despite this, domestic GPU start-ups in China have made significant progress in hardware and software development in recent years. However, CAICT highlighted the engineering challenges involved in transitioning models trained on Nvidia GPUs to domestic solutions.
Due to export restrictions, the A100 and H100 GPUs, crucial for AI model training and deployment, have been modified by Nvidia to create the A800 and H800. Nevertheless, these modified chips were also barred from being sold to China, complicating the situation further.