DeepSeek masters PTX and may move to Chinese GPU
China AI Lab DeepSeek, which has recently rocked US financial markets, may begin to use domestic graphic processors to train their models. According to the report Hong Kong press, the developers of DeepSeek have expertise in low-level programming in the language Parallel Thread Execution (PTX)which allows them bypass NVIDIA restrictions and effectively use even less powerful chips.
DeepSeek is currently using American H800 on Hopper architecture, produced by 5nm process technology. However, given US sanctions, the company may begin the transition to Chinese GPU, despite their limited performance. PTX allows deeper control over the chip's operation, which makes it possible to compensate for the lag of Chinese processors due to optimized program code.
Experts believe that DeepSeek can adapt NVIDIA drivers to work with SMIC chips — China's largest semiconductor manufacturer. These GPU will be created according to 7nm process technology, which is inferior to American analogues, but optimization through PTX can bridge the gap. If Chinese engineers can fully integrate PTX, it will become critical breakthrough for China's AI industry, paving the way for independence from NVIDIA and AMD.