Archive — US — turtlesai
MUSA SDK: China Responds to CUDA with Its Own Domestic AI Platform
Moore Threads launches MUSA SDK 4.0.1, a domestic alternative to NVIDIA's CUDA, compatible with various CPUs and designed for AI, streamlining code migration for developers. The SDK includes libraries for mathematical operations and supports PyTorch for running AI models on Moore Threads GPUs, reinforcing China's position in the GPU and AI landscape.
Key points
- MUSA SDK 4.0.1 introduced by Moore Threads as an alternative to CUDA.
- Supports Intel, ARM, and Chinese CPUs (Hygon, Kylin, Loongson), enhancing hardware compatibility.
- MUSIFY tool allows easy migration of CUDA code to MUSA, simplifying development.
- Libraries like muBLAS, muFFT, and muThrust provide necessary support for mathematical and parallel processing.
- torch_musa enables PyTorch integration, facilitating AI model execution on Moore Threads GPUs.
- Founded by Zhang Jianzhong, a former NVIDIA executive, Moore Threads aims to dominate domestic AI and GPU markets.
Why it matters
- Reflects the growing trend of domestic technology development in China, especially in AI and computing.
- Aligns with the global shift towards decentralization of AI processing capabilities away from dominant players like NVIDIA.
- Indicators of rising geopolitical tensions affecting tech supply chains and innovation strategies in response to external dependencies.
- Part of broader 2025 IT trends focusing on increased competition in AI hardware markets and development of alternative platforms.
The launch of MUSA SDK 4.0.1 positions Moore Threads as a significant player in the AI space, aiming to reduce dependence on foreign technologies and foster innovation within China.