Moore Threads Launches MUSA SDK: China’s In-House Alternative to NVIDIA’s CUDA

Moore Threads, a prominent Chinese technology company, is actively seeking its foothold in the competitive artificial intelligence (AI) landscape. The firm has recently enhanced its MUSA SDK, a notable alternative to NVIDIA’s well-established CUDA framework.

Enhanced MUSA SDK: Compatibility with Intel & ARM Processors and Code Porting from CUDA

NVIDIA’s CUDA ecosystem has long dominated the software landscape, celebrated for its robust support system and frequent updates that leave little room for competitors. However, due to shifting geopolitical dynamics, Chinese tech firms have found it increasingly challenging to depend exclusively on NVIDIA products. In response to these conditions, Moore Threads has introduced significant upgrades to its MUSA SDK, aiming for wider implementation among developers.

The MUSA SDK is tailored specifically for use with Moore Threads’ GPUs, providing capabilities for parallel computing and handling AI tasks across multiple platforms. This software suite comprises runtime libraries, drivers, and instruction sets, all designed to facilitate integration with existing systems. The SDK is structured into several components that include dedicated toolkits and libraries geared toward specific applications, which we will explore further below.

A notable feature of the MUSA SDK is the inclusion of a powerful tool called MUSIFY, which enables developers to easily port existing CUDA-based code to the MUSA environment. This functionality promotes smoother transitions for those adopting the new technology. Additionally, the SDK is equipped with specialized libraries such as muBLAS, muFFT, and muThrust, which are optimized for applications requiring advanced mathematical computations and accelerated processing capabilities.

With the recent launch of version 4.0.1 of the MUSA SDK, support has been extended to Intel processors, as well as domestic CPUs from brands like Hygon, Kylin, and Loongson, catering to standard workload requirements. While Moore Threads may not yet match NVIDIA’s CUDA capabilities, providing a local alternative allows smaller-scale developers to leverage in-house GPUs at a more economical rate, fostering a sense of technological independence and national pride.

Source & Images