MUSA vs CUDA: China Builds Its Own GPU Ecosystem

Moore Threads has rolled out an updated version of its MUSA SDK — a software platform designed for GPU computing. The move is part of China’s broader push to develop self-reliant tech amid U.S. export bans on high-end chips, semiconductor manufacturing tools, and HBM memory.

Positioned as a homegrown alternative to NVIDIA’s CUDA, MUSA SDK supports both Moore Threads’ own MTT-series GPUs and Intel CPUs, running on Ubuntu and Kylin OS. The platform includes a compiler, libraries for machine learning and mathematical operations, and a suite of debugging tools. Special emphasis is placed on compatibility with existing CUDA projects — the included Musify tool automatically converts CUDA code to run on the MUSA architecture.

The developer’s website features demo projects showcasing neural network image generation, 3D face modeling, and audio processing. Some of these examples run on the server-grade MTT S3000 GPU, highlighting the platform’s aim at not just research, but real-world industrial use as well.

Analysts say MUSA still lags behind CUDA in terms of optimization and developer support, but its evolution could satisfy local market needs — especially for small and mid-sized operations. In the long run, the success of the ecosystem will hinge on the performance of Chinese GPUs and the continued expansion of the SDK’s capabilities.

Other CUDA alternatives, like AMD’s ROCm, have also struggled due to limited hardware support. Moore Threads is betting on a tightly controlled ecosystem — a strategy that may boost compatibility with domestic tech, but could also limit adoption outside of China.