China’s AI lab DeepSeek, whose AI models wreaked havoc on Wall Street last week, can use Chinese GPUs for future models, suggests a report. DeepSeek’s ability to purportedly reduce costs significantly over American models while using GPUs from chip giant NVIDIA Corporation upended markets as investors wondered whether billions of dollars in capital expenditure for AI development were needed. At the heart of the firm’s model development is its ability to use low-level programming language on NVIDIA’s GPUs called Parallel Thread Execution (PTX) instead of NVIDIA’s offered language called Compute Unified Device Architecture (CUDA).
DeepSeek Could Rely On PTX Expertise To Eke Out Greater Performance From Domestic Chinese GPUs, Says Report
Chinese semiconductor fabrication has been a frequent target of US sanctions which have prevented its largest foundry, the Semiconductor Manufacturing International Corporation (SMIC) from acquiring the latest chip making machines from Dutch giant ASML. These machines are indispensable for making the most advanced chips, with several reports suggesting that SMIC is restricted to manufacturing chips via the older 7-nanometer process technology family.
A key factor in DeepSeek’s ability to match performance of some Western AI models while using fewer and less performant NVIDIA GPUs has been the firm’s ability to use a programming language called Parallel Thread Execution or PTX. PTX allows engineers to control a chip’s basic functions more deeply than other languages, and there are relatively fewer engineers that are capable of using it.
DeepSeek’s decision to use the notoriously hard PTX stemmed from US GPU sanctions as the firm was able to secure either limited quantities of GPUs or only those chips that had lower performance. However, its expertise in using PTX can allow the firm to rely on domestic Chinese GPUs for future models suggests a report from the Hong Kong press.
The report believes that since PTX allows programmers to tightly control a GPU’s operations, it will enable DeepSeek’s engineers to extract more performance from domestic Chinese GPUs. NVIDIA’s Hopper H800 GPUs that DeepSeek claims to use are built using the 5-nanometer manufacturing process node. Launched in 2023, they had used the latest manufacturing process technology in the industry at the time.
The report adds that by understanding NVIDIA’s PTX implementation and how drivers control a chip’s functions, DeepSeek can use similar programs with the Chinese GPUs. Any Chinese GPUs will likely be built with older manufacturing process technologies and offer significantly lower performance than NVIDIA’s latest Blackwell GPUs.
