Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Chinese companies are resorting to chips repurposed from standard PC gaming products to develop artificial intelligence tools, after Washington blocked US exports of high-performance processors.
Thousands of Nvidia gaming graphics cards are being stripped of their core components in factories and workshops every month, before being installed on new circuit boards, according to two factory managers and two chip buyers familiar with the situation.
Industry experts said the repurposing of chips from cards designed to be slotted into consumer PC motherboards to improve gaming graphics amounted to a rough workaround to tackle the lack of high-end processors in China.
While Nvidia’s gamer-focused products have raw computing power, they are not as capable in the high-precision calculations needed for training some large language models with bigger data sets. Due to the limits of interconnection speeds between chips, it is also challenging to overcome this by simply grouping more chips in computing clusters.
“This is a desperate move by Chinese companies under the export controls. Just like using a kitchen knife to create artwork, it’s doable, but the effect is suboptimal,” said Charlie Chai, an analyst at research group 86Research.
The Biden administration tightened export controls for cutting-edge AI chips in October, making it more difficult for chip companies such as Nvidia to sell high-performance semiconductors to China.
Demand for the graphics processing units sourced from the gaming cards has surged in the past month, according to people familiar with the situation. One of the factory managers said workers had disassembled more than 4,000 Nvidia gaming cards in December, more than four times the number in November.
Customers for the repurposed components have been mainly public enterprises and small AI labs, which had not stockpiled enough Nvidia server chips before the new US export controls took effect, according to the two factory managers.
They declined to reveal more details about their clients due to the sensitivity of the issue.
However, industry experts and analysts warned that modifications to Nvidia’s products violated the company’s intellectual property rights, while some of the gaming cards could be banned from being sold to China at any time.
Nvidia’s most powerful gaming graphics board, the GeForce RTX 4090, was one of the most popular models to be repurposed, but it has now been blocked from being sold to China, the company said.
To comply with the latest controls, Nvidia came out with a slower version of the banned cards last month, called the GeForce RTX 4090 D, which is 5 per cent slower than versions sold outside China, according to the company.
One of the factory managers said the performance gap between the modified versions of the 4090 D and 4090 would be “more significant”, which could mean the slower version was not powerful enough for large language model training. The manager said a batch had been procured for further verification.
Nvidia told the Financial Times that “disassembling gaming cards is not a viable way to create data centre compute clusters for AI”, adding that gaming products were “designed, manufactured, and marketed for individual gamers and consumers”.
Nvidia has developed three chips tailored for China that meet the region’s growing demand for AI systems while complying with US export controls, enabling the company to maintain its foothold in one of its most important markets.
However, the performance of these chips is substantially weaker than those previously sold in China, and they will not be widely available until March, according to three people familiar with the situation.
Chinese customers have also objected to the prices Nvidia hopes to set for these inferior processors, which are close to those of their more powerful banned counterparts, the three people said.
However, the options for migrating to China’s alternative chip ecosystem under development are limited, leading some companies to turn to Nvidia’s less expensive gaming chips.
“We don’t know whether such a reinvention will be successful, but we hope that these machines will be usable, at least in the short term,” said a buyer.
Read the full article here