Despite last year's US microchip export controls aimed at stalling China's advancement in supercomputers and AI systems, China's tech sector has experienced minimal disruption. The restrictions targeted shipments of Nvidia Corp (NVDA.O) and Advanced Micro Devices Inc (AMD.O) chips, which are essential for developing AI systems and chatbots.
Nvidia has developed variants of its chips for the Chinese market that comply with US regulations. According to industry experts, the latest chip, the Nvidia H800, may require 10% to 30% more time to perform certain AI tasks and could potentially double some expenses compared to the fastest US chips.
Nevertheless, the slower Nvidia chips are an upgrade for Chinese companies. Tencent Holdings (0700.HK), a leading Chinese tech firm, estimated in April that using Nvidia's H800 would reduce the training time for its largest AI system by over half, from 11 days to four days.
Charlie Chai, a Shanghai-based analyst with 86Research, stated, "The AI companies that we talk to seem to see the handicap as relatively small and manageable." This situation exposes the US challenge of impeding China's progress in high tech without adversely affecting US companies.
The export restrictions comprise two parts. The first imposes a limit on a chip's ability to calculate extremely precise numbers, effectively curbing supercomputers used in military research. The second restriction is on chip-to-chip transfer speeds, which impacts AI.
Although Nvidia has not revealed the China-only H800 chip's performance specifics, a specification sheet viewed by Reuters displays a chip-to-chip speed of 400 gigabytes per second, less than half of the 900 gigabytes per second peak speed for Nvidia's flagship H100 chip available outside China.
Naveen Rao, CEO of MosaicML, a startup that focuses on optimizing AI models to run efficiently on limited hardware, estimated a 10-30% system slowdown. However, he believes this limitation is manageable, saying, "There are ways to get around all this algorithmically. I don't see this being a boundary for a very long time - like 10 years."
Furthermore, AI researchers are attempting to streamline their systems to reduce training costs for AI processes. These streamlined models will require fewer chips, ultimately diminishing the effect of US speed limitations.
Cade Daniel, a software engineer at Anyscale, a San Francisco startup that offers software to help companies perform AI work, commented, "This export restriction is noticeable, but it's not quite as devastating as it could have been."
