AMD GPUs Now ‘Indirectly” Support NVIDIA CUDA Libraries With ROCm Using ZLUDA

AMD GPU owners can now effortlessly run CUDA libraries and apps within ROCm through the use of ZLUDA, an Open-Source library that effectively ports NVIDIA CUDA apps over to ROCm that does not require code adaption.

The practice of "transitioning" codebases from one project/language to the other is common since it allows developers to code to access a wider "arsenal" of features and makes maintenance easier. In this case, however, an interesting development has occurred since Team Red has apparently hastened its efforts of porting the CUDA platform to run it with AMD's ROCm but contacted a developer who has already been involved in a similar implementation.

Phoronix reports that AMD's ROCm magic now supports CUDA APIs in a "workaround" method, which involves utilizing ZLUDA, an open-source porting project designed originally to facilitate NVIDIA CUDA support on Intel GPUs. While ZLUDA offered intriguing possibilities in the domain of code porting, it faced technical and practical challenges that ultimately halted its development, especially when it came to the interest of companies such as Intel at that time. However, the developer behind it, Andrzej Janik, was contacted by AMD for a pretty exciting experiment back in 2022, the results of which are now finally evident.

Based on extensive testing by Phoronix, it is disclosed that ZLUDA "almost" runs perfectly with AMD's Radeon GPUs on with ROCm utilizing NVIDIA's CUDA-loaded libraries. Moreover, it didn't require any changes to the existing code either. Despite the project's development getting halted due to AMD's apparent backout, the developer was kind enough to open-source the work after he met the deadline, which made it possible for Phoronix to test out to see what it brought onboard.

Since code porting opened up new possibilities, Phoronix has managed to run Blender 4.0 with CUDA libraries, and surprisingly enough, the testing results show that NVIDIA and AMD are head-to-head in terms of rendering performance. Moreover, the developer Andrzej Janik himself tested the performance of Radeon GPUs with CUDA-backing through synthetic benchmarking, and the figures obtained were interesting to witness.

I believe ZLUDA might have huge potential moving forward since it is a platform that bridges ROCm and CUDA stacks rather than separates them out, allowing developers to leverage specific capabilities from both platforms. Of course, we can't see native CUDA support on Radeon GPUs, but such resources could prove to be very useful as we advance into the era of AI, and the developer himself is optimistic about the future of ZLUDA, claiming that he is testing out NVIDIA's upscaling capabilities with ZLUDA on RDNA GPUs.

News Source: Phoronix