Chinese companies are resorting to repurposing Nvidia's gaming graphics cards for their AI models in a move described as "desperate" to circumvent U.S. export sanctions, according to reports from the Financial Times (FT). The workaround involves stripping components from Nvidia gaming chips and installing them onto new circuit boards, forming makeshift clusters to address the restrictions imposed by the Biden administration.
While Nvidia's gaming chips are powerful, they lack the precision required for high-precision calculations used in training large language models with extensive datasets. 86Research analyst Charlie Chai likened the approach to using a kitchen knife for artwork, acknowledging its feasibility but noting suboptimal results.
Despite the challenges, demand for Nvidia's gaming GPUs has surged, with one factory reportedly disassembling over 4,000 Nvidia chips in December alone—more than quadruple the number in November. The Biden administration's restrictions on exporting AI chips to China have particularly impacted Nvidia, a dominant player in the market.
In response to export restrictions, Nvidia had previously developed hardware tailored for local markets, such as the slower 800-series chip. However, these China-specific chips were also subject to updated export restrictions introduced in October. Consequently, U.S. chipmakers are limited to exporting slower systems to China, with Nvidia unveiling three compliant units last November.
The repurposing of Nvidia gaming graphics chips continues despite potential intellectual property violations, with public enterprises and small AI labs in China reportedly among the purchasers. The popular GeForce RTX 4090 graphics cards have been a focus of repurposing efforts, leading to the creation of a market-specific version, the GeForce RTX 4090 D, to comply with export regulations.
Nvidia emphasized that dismantling gaming chips is not a viable method for creating data center compute clusters for AI, emphasizing that these chips are designed for individual gamers and consumers. The FT's sources suggest that the China-specific 4090 D version may not be powerful enough for language model training, raising questions about the sustainability of the current workaround.