The U.S. microchip export restrictions put in place last year to halt China’s development of supercomputers used to create nuclear bombs and ChatGPT are barely making a dent in the country’s IT industry.

The regulations banned shipments of Nvidia Corp. and Advanced Micro Devices Inc. chips, which have evolved into the benchmark for building chatbots and other AI systems across the global technology sector.

However, Nvidia has developed slower versions of its chips for the Chinese market in order to comply with American regulations. According to industry insiders, the Nvidia H800, which was introduced in March, will likely take 10% to 30% slower to do various AI tasks and may cost twice as much as Nvidia’s top U.S. chips.

For Chinese companies, even the slower Nvidia chips are an improvement. One of the biggest digital companies in China, Tencent Holdings, said in April that systems utilising Nvidia’s H800 will reduce the time it takes to train its largest AI system by more than half, from 11 days to four days.

“The AI companies that we talk to seem to see the handicap as relatively small and manageable,” said Charlie Chai, a Shanghai-based analyst with 86Research.

The back-and-forth between government and business reveals how difficult it is for the United States to restrict China’s advancement in high technology without harming American businesses.

The United States’ goal in establishing the limits was in part to prevent a shock that would cause the Chinese to completely stop using American chips and step up their own chip-development efforts.

“They had to draw the line somewhere, and wherever they drew it, they were going to run into the challenge of how to not be immediately disruptive, but how to also over time degrade China’s capability,” said one chip industry executive who requested anonymity to talk about private discussions with regulators.

There are two aspects to the export limitations. The first imposes a cap on a chip’s capacity to compute extremely accurate numbers, a move intended to restrict the use of supercomputers for military research. Sources in the chip sector stated that was a successful move.

However, in AI tasks like complex language models, where the amount of data the chip can process is more crucial, calculating extremely precise numbers is less important.

Although it has not yet begun shipping the chips in large quantities, Nvidia is selling the H800 to China’s largest technology companies, including Tencent, Alibaba Group Holding Ltd, and Baidu Inc  for use in such work.

“The government isn’t seeking to harm competition or U.S. industry, and allows U.S. firms to supply products for commercial activities, such as providing cloud services for consumers,” Nvidia said in a statement last week.

According to the business, China is a significant market for American technology companies, and the sale of goods there helps both Nvidia and its American partners create jobs.

“The October export controls require that we create products with an expanding gap between the two markets,” Nvidia said last week. “We comply with the regulation while offering as-competitive-as-possible products in each market.”

In a different statement this week, Bill Dally, the chief scientist of Nvidia, predicted that “this gap will grow quickly over time as training requirements continue to double every six to 12 months.”

The U.S. Commerce Department’s Bureau of Industry and Security, which is in charge of enforcing the regulations, did not respond to a request for comment.

AI is impacted by the second U.S. restriction on chip-to-chip transmission speeds. The models that underlie innovations like ChatGPT are too big to fit on a single chip. Instead, they must be dispersed among a large number of chips—often thousands at once—which must all interact with one another.

Performance information for the Nvidia H800 chip, which is only accessible in China, has not been released by the company, but a specification sheet obtained by Reuters reveals a chip-to-chip speed of 400 GB/s, which is less than half the peak speed of 900 GB/s for Nvidia’s flagship H100 chip, which is sold outside of China.

That speed, according to some in the AI industry, is still sufficient. A 10–30% system delay was reported by Naveen Rao, CEO of a business called MosaicML that specialises in assisting AI models to perform better on constrained hardware.

“There are ways to get around all this algorithmically,” he said. “I don’t see this being a boundary for a very long time – like 10 years.”

Money is beneficial. A chip made in China that completes an AI training task twice as quickly as a chip made in the United States can nonetheless complete the task.

“At that point, you’ve got to spend $20 million instead of $10 million to train it,” said one industry source who requested anonymity because of agreements with partners. “Does that suck? Yes it does. But does that mean this is impossible for Alibaba or Baidu? No, that’s not a problem.”

Additionally, AI researchers are working to reduce the size of the enormous systems they have constructed in order to lower the cost of developing products like ChatGPT and other procedures. These will use fewer chips, which will lessen chip-to-chip communications and the effect of US speed limits.

According to Cade Daniel, a software engineer at Anyscale, a San Francisco startup that offers software to enable companies execute AI work, the industry believed two years ago that AI models would continue to grow in size.

“If that were still true today, this export restriction would have a lot more impact,” Daniel said. “This export restriction is noticeable, but it’s not quite as devastating as it could have been.”

(Adapted from Reuters.com)

Categories: Creativity, Economy & Finance, Entrepreneurship, Geopolitics, Regulations & Legal, Strategy