Why China's AI models are still trained on NVIDIA chips
What's the story
China's leading artificial intelligence (AI) models are still being trained using NVIDIA chips, as per industry insiders speaking to South China Morning Post. The high cost of switching to local semiconductors is a major roadblock in Beijing's quest for self-sufficiency. Despite advancements in domestic hardware, the transition to new chip architecture is proving to be an engineering challenge.
Technical challenge
Software ecosystem poses challenge in transition
The software ecosystem is another major hurdle in the transition from NVIDIA to local semiconductors.
The tech giant's Compute Unified Device Architecture (CUDA) has been the go-to platform for AI development.
In contrast, Huawei's alternative, Compute Architecture for Neural Networks (CANN), demands developers to rewrite and optimize large amounts of code, making the transition even more difficult.
Cost implications
Transition could increase time and costs by 50%: Developer
James Wang, an AI model developer at a Shanghai-based university research institute, said their existing training pipelines are dependent on CUDA.
He added that "CUDA code cannot run directly on Ascend and requires extensive rewriting."
Wang estimated that switching existing workflows to Huawei's Ascend chips could increase time and costs for his team by at least 50%.