Microsoft is Developing a Tool to Run NVIDIA CUDA Code on AMD GPUs

The hardware market has existed for years under conditions of a tight lock-in to a single manufacturer. NVIDIA's CUDA software platform has become the de facto standard, forcing developers to use exclusively this vendor's graphics processors, while attempts to create universal alternatives have so far not yielded significant success.

This situation could change thanks to the work of Microsoft's specialists. According to available information, the company is actively developing tools to adapt code written for CUDA to run on AMD's open ROCm platform. This would allow existing AI models to run on more affordable AMD chips without a complete rewrite.

A similar working principle was previously demonstrated by the ZLUDA project. Microsoft's approach, it seems, also involves converting CUDA commands into instructions understandable by ROCm on the fly. The main economic sense behind this initiative is a shift in demand within the AI sphere. While costly computational power for training neural networks was previously key, the stage of inference—the practical application of ready-made models—is now coming to the fore.

Deploying these models doesn't always require top-tier, expensive NVIDIA solutions, which makes the arguments in favor of AMD increasingly compelling. However, experts estimate that ROCm software still lags behind the CUDA ecosystem in terms of polish and maturity, and in some scenarios, this can lead to a loss of performance, a critical factor for commercial data centers.

Increased competition between chip manufacturers and reduced dependence on a single technology platform could, in the long term, stimulate innovation and curb the growth of graphics card prices. Furthermore, conversion tools could make life easier for developers using technologies like ray tracing or AI upscaling, allowing them to adapt these features for AMD graphics cards more quickly.