ZLUDA attacks with renewed strength

The latest iteration will focus on CUDA-based AI software.

At the beginning of the year, we reported on the third version of the ZLUDA project, which runs binary CUDA files, which made it possible to run certain CUDA programs on AMD GPUs, but in the end AMD did not ask for it, so they asked the developer to remove the project. Andrzej Janik complied with this, but he already indicated that ZLUDA would live on in some form.

Now it’s time to come back, because the new edition of ZLUDA has become available, which is on GitHub below accessible. For now, it’s still rudimentary code, but the focus is on CUDA-based AI software (such as Llama.cpp, PyTorch, TensorFlow), and the project is trying to run them on GPUs from multiple manufacturers.

Initially, the priority is to support AMD’s RDNA GPUs on the ROCm 6.1 or later platform, which is needed to have a basic working codebase that can later be ported to other GPUs such as Intel. This process can be facilitated by keeping the entire project open source, so anyone can contribute in any form.

The question may arise as to whether AMD can cancel this project as well, since the previous one was about the same thing. Probably not. The problem with the previous codebase could be that AMD probably wrote a lot of code in it, which in the end they didn’t want Andrzej Janik to use. However, if ZLUDA is rebuilt, it will be the project’s own code, which AMD doesn’t really have much to do with, even ROCm can be used quite freely. The only thing that was important to AMD here was that if NVIDIA might not agree with ZLUDA’s goal and they took the case to court, then AMD should not be targeted with it.