China is exploring new approaches to reduce its reliance on Nvidia’s CUDA software, which plays a key role in the company’s dominance in artificial intelligence.
Wei Shaojun, an executive at the China Semiconductor Industry Association, has advised the domestic AI industry to develop alternatives to Western technologies, including CUDA.
CUDA Seen as Nvidia’s Key Advantage
Nvidia CEO Jensen Huang has repeatedly described CUDA as the company’s strongest advantage, highlighting the importance of its software ecosystem in driving adoption.
Developers widely use CUDA because it is mature and well-supported, which also ties them to Nvidia’s hardware.
Proposal Focuses on Software-Defined Chips
Instead of building a direct alternative to CUDA, Wei Shaojun suggested a different approach known as software-defined chips (SDCs).
This concept shifts more of the computing logic into software rather than relying on fixed hardware designs.
How Software-Defined Chips Work
In an SDC system, developers do not need a CUDA-like layer to run workloads. Instead, chips use a flexible grid that is configured through instructions generated by a compiler.
This means the software and code are not tied to a specific hardware instruction set, making the system more adaptable.
Unlike traditional GPUs, which rely on schedulers to manage tasks, SDCs use deterministic compilation, where data movement is planned and controlled in advance down to precise timing.
Challenges and Trade-Offs
Wei Shaojun noted that building alternatives to CUDA through translation layers and separate ecosystems would require significant resources.
He described the software-defined chip approach as a more practical option, although it comes with challenges. These include complex compiler requirements, routing and branching issues, and design changes that differ from traditional hardware models.
Existing Examples and Industry Context
Some existing technologies follow similar principles, including systems developed by companies like SambaNova Systems and Groq.
However, these solutions are typically designed for specific workloads and are not direct replacements for GPUs.
