May 3 (Reuters) – U.S. microchip export restrictions imposed last year to freeze the development of China’s supercomputers used to develop nuclear weapons and artificial intelligence systems like ChatGPT have been imposed on China’s technology sector. has had minimal impact on
The rule restricts shipments of chips from Nvidia Corp (NVDA.O) and Advanced Micro Devices Inc (AMD.O), which have become the global tech industry standard for developing chatbots and other AI systems. doing.
However, Nvidia has created a variant of the chip for the Chinese market that has been slowed down to match US regulations. Industry experts told Reuters that the latest Nvidia H800, announced in March, will likely take 10% to 30% longer to perform some AI tasks, making it the fastest U.S.-made chip from Nvidia. He said the cost could double compared to .
Even the slowdown in Nvidia chips shows improvement for the Chinese company. Tencent Holdings (0700.HK), one of China’s biggest tech companies, said in April that his H800-powered system from Nvidia cut the time it takes to train the company’s largest AI system in half from 11 days to 4 days. I assumed it would be shorter.
Shanghai-based 86Research analyst Charlie Chai said:
The back-and-forth between government and industry exposes the U.S. challenge to slow China’s progress in high-tech without hurting U.S. companies.
Part of the U.S. strategy in setting the rules was to avoid shocking China to the point of ditching U.S. chips altogether and doubling down on its own chip development efforts.
“They have to draw the line somewhere, and wherever they do, they’re going to face the challenge of how to degrade China’s capabilities over time, rather than creating chaos right away. No,” said a chip industry executive who wished to remain anonymous. To discuss private discussions with regulators.
There are two parts to export restrictions. The first caps the chip’s ability to compute highly accurate numbers, a measure designed to limit the supercomputers available for military research. Chip industry sources say it was an effective move.
But computing highly accurate numbers is less relevant for AI tasks like large-scale language models, where the amount of data a chip can handle is more important.
Nvidia has sold the H800 to some of China’s biggest tech companies, including Tencent, Alibaba Group Holding (9988.HK) and Baidu (9888.HK), but has yet to start mass shipping the chips.
In a statement last week, Nvidia said, “The government does not seek to harm competition or U.S. industry and does not allow U.S. companies to supply products for commercial activities, such as providing cloud services to consumers. There are,” he said.
China is an important market for U.S. tech companies, and selling products in China will create jobs for both Nvidia and its U.S.-based partners, the company added.
Nvidia said last week, “October export controls require us to make products that widen the gap between our two markets. To do.”
“This gap will grow rapidly over time as training requirements continue to double every 6 to 12 months,” Nvidia chief scientist Bill Dally said in a separate statement this week. ‘ said.
A spokeswoman for the Bureau of Industry and Security, the division of the U.S. Department of Commerce that oversees the regulation, did not return requests for comment.
It doesn’t stop even if it’s late
A second limitation in the United States concerns chip-to-chip transfer speeds that affect AI. The model behind technology such as ChatGPT is too big for him to fit on one chip. Instead, they have to be distributed over many chips (sometimes thousands at a time), all of which need to communicate with each other.
Nvidia hasn’t revealed performance details for its China-only H800 chips, but specs seen by Reuters show chip-to-chip speeds of 400 gigabytes per second, the highest speed of Nvidia’s flagship H100 chips. is less than half of 900 gigabytes per second. Available outside China.
Some in the AI industry still think it’s fast enough. Naveen Rao, CEO of a startup called MosaicML, said he specializes in helping AI models run better on limited hardware, and his system speed is 10-10%. We estimate a 30% drop.
“There are ways around all of this algorithmically,” he said. “I don’t think this will be the boundary for very long – ten years or so.”
Money helps. A Chinese chip that takes twice as long to complete an AI training task can complete the task better than a faster US chip.
“At that point, you have to spend $20 million on training instead of $10 million,” said an industry source who requested anonymity because of the deal with partners. “Is that terrible? Yes it is. But does that mean this is not possible for Alibaba or Baidu? No, it doesn’t matter.”
Additionally, AI researchers are looking to slim down the large systems they have built to reduce the cost of training products like ChatGPT and other processes. These require fewer chips, reduce chip-to-chip communication, and reduce the impact of US speed limits.
Two years ago, the industry thought AI models were getting bigger and bigger, says Cade Daniel, software engineer at Anyscale.
“If that were true today, this export restriction would have a much bigger impact,” Daniel said.
Reporting by Stephen Nellis and Jane Lee from San Francisco and Josh Ye from Hong Kong Editing by Peter Henderson and Matthew Lewis
Our standards: Thomson Reuters Trust Principles.
