Why China Recycles Gaming Chips to Train AI

In recent months, China has faced a significant lack of processor high end due to restrictions imposed by the United States on the export of chips for artificial intelligence. In response to this situation is practice recycle Nvidia graphics card chips and dismantle them from gaming PCs. An inefficient solution due to a difficult situation, given that although Nvidia products designed for video games, such as the GeForce RTX 4090, have very high processing power, they are not effective in high precision calculations needed to train some large language models.

The Chinese factory director said so Financial Times that its workers dismantled more than 4000 Nvidia gaming cards in December, which is more than four times the number in November. According to the English newspaper, in addition, the greatest interest of customers in redesigned parts was manifested above all public companies and small artificial intelligence labswhich didn’t have enough stock of Nvidia chips before the new US export controls went into effect.

USA stop

The update to US restrictions in October, strongly supported by the administration led by President Joe Biden, has a clear objective: limit China’s access to advanced chips that “could feed discoveries in the field of artificial intelligence and in sophisticated computers” explained the US Secretary of Commerce Gina Raimondo. From 2022 The US imposes restrictions on the performance and speed of chips exportable to Asia, making it difficult for American companies to sell its high-end chips. Nvidia suffered a strong stock market depreciation in the last quarter of 2023, so it decided to respond to Washington’s sanctions by submitting less powerful versions of their cards that are not subject to the export ban. The distribution of the three new low-power chips (H20, L20 and L2) adapted for the Chinese market will take some time (the H20 was also delayed until March), but in any case they do not seem to have convinced completely. The Chinese market looking for high-end chips for applications in technologies using artificial intelligence.

Top cloud service providers in China such as Alibaba and Tencentare testing Nvidia chips introduced in November 2023. But American companies have already said that will order fewer chips in 2024 than originally planned, because they prefer more powerful versions. According to The Wall Street Journalthe performance advantage of Nvidia’s “downgraded” chips over domestic Chinese alternatives is diminishing in the near future, making locally produced chips more attractive to buyers.

Despite Nvidia currently holds about 80% of China’s high-end AI chip marketthis percentage could reduce to 50-60% over the next five years. Outside Huaweithan with his Exit 910B It is considered by many to be the Chinese company making the most progress in AI chips as well Baidu is moving in this direction. In 2020, it launched a long-term plan to create chips: from the first model, Kunlunxin, the second version was developed in 2021 and the third is expected in 2024. After the acquisition of Zhongtian Micro Systems and the establishment of T-Head Semiconductor Alibaba has started making its own AI chips: the most powerful isHanguang 800.

Leave a Comment