The Day After: US Restricts AI Chips to China
- By Paul Mah
- September 07, 2022
According to various reports this week, U.S. government officials have ordered Nvidia to stop exporting its A100 and H100 GPUs to China. Similarly, AMD has also said that new licensing requirements now prevent the export of its latest MI250 AI accelerator.
For those not already aware, Nvidia’s upcoming H100 is slated to compete directly with the MI250X, while the older A100 is Nvidia’s mainstay of enterprise GPU and is commonly deployed in supercomputers.
For instance, Thailand’s National Science and Technology Development Agency's (NSTDA) supercomputer uses 704 Nvidia A100 Tensor Core GPUs.
For its part, the new Instinct MI200 family of accelerators from AMD was unveiled in November last year and represents a major upgrade over the previous Instinct MI100.
Indeed, the highest-end MI250X offering provides almost five times faster performance than Nvidia’s A100 GPU for double precision (FP64) HPC applications for more than 380 teraflops of peak theoretical half-precision performance.
Major escalation
While the move will have few visible impacts at the moment, the development signals a major escalation of a U.S. campaign to stymie China’s technological capability, according to a report by Reuters.
It also comes mere weeks after the U.S. banned the sale of electronic design automation (EDA) software used to design high-performance chips, as well as moves to fund chip facilities in the U.S. through the US Chips and Science Act.
Nvidia has an 80.6 percent share in the global market for artificial intelligence processors used in the cloud and data centers in 2020, says Omdia. And according to TechCrunch, China made up 26% of Nvidia’s revenues in 2021. It was no surprise that the stock market prices of both AMD and Nvidia fell with the news.
While the opinion differs as to whether China can overcome the hurdles designed to hurt its AI ambitions, the biggest immediate losers would undoubtedly be China’s Internet giants, cloud providers, and AI firms reliant on high performance AI hardware from Nvidia or AMD.
But beyond dealing a heavy blow to China’s AI sector, the bigger implications of the ban will be its impact on the next generation of AI applications.
Creating a schism
According to McKinsey, China produced about one-third of both AI journal papers and AI citations worldwide in 2021, and the country attracts some USD17 billion for AI start-ups.
Overall AI adoption in China is also high, with China being one of the first countries to publish ethical guidelines governing AI algorithms to guide the use of AI technology. In addition, AI is also expected to create upwards of USD600 billion in economic value annually.
In a nutshell, China is a juggernaut in AI development and adoption.
So, what can the AI industry in China do now? With no ability to access AI processors from AMD or Nvidia, China’s internet giants and cloud providers will now be forced to either significantly boost their investments on their own AI chips or place their bets on China-made AI offerings.
And with so much at stake, this amount will hardly be trivial.
To be clear, switching to another AI platform will be a herculean task that will take years of sustained effort. This is due to the dominance of Nvidia’s CUDA parallel computing platform and APIs. Likewise, existing code and AI libraries written for AI hardware from AMD or Nvidia will need to be rewritten.
But much like how Huawei was forced to eschew Google’s Android platform for its own Harmony operating system, survival dictates that the same will have to happen with China’s mammoth AI industry.
And once that happens, there is no turning back.
Paul Mah is the editor of DSAITrends. A former system administrator, programmer, and IT lecturer, he enjoys writing both code and prose. You can reach him at [email protected].
Image credit: iStockphoto/metamorworks
Paul Mah
Paul Mah is the editor of DSAITrends, where he report on the latest developments in data science and AI. A former system administrator, programmer, and IT lecturer, he enjoys writing both code and prose.