In the evolving landscape of GPU computing, a project by the name of "ZLUDA" has managed to make Nvidia's CUDA compatible with AMD GPUs. Historically, CUDA, a parallel computing platform and programming model developed by Nvidia, has been exclusively available for Nvidia GPUs. This exclusivity has posed challenges for developers and researchers who wish to leverage CUDA's powerful computing capabilities on non-NVIDIA hardware. There are many popular CUDA-powered programs out there, including PyTorch and Matlab.

ZLUDA, initially inspired by efforts to adapt CUDA applications for Intel graphics through a similar project of the same name, aims to bridge this gap for AMD GPUs. By translating CUDA calls into something that AMD's ROCm (Radeon Open Compute) platform can understand, ZLUDA enables CUDA applications to run on AMD hardware with minimal to no modifications required from the developers. The project was undertaken by Andrzej Janik, who reached out to Phoronix and shared his ZLUDA implementation for AMD ROCm.

This project was initially funded by AMD, and is the product of two years of work in making it compatible with CUDA. AMD discontinued funding it, but a stipulation of the contract was that the project could be open-sourced and released by Janik once the contract ended. That's why we're getting to see it now, and apparently the performance is impressive. As it stands, Janik makes it clear that this is likely the end of the project, stating that "it's now abandoned and will only possibly receive updates to run workloads I am personally interested in (DLSS)" and that "we've run out of GPU companies" given that both AMD and Intel are now uninterersted.

As Phoronix reports, CUDA-enabled software runs without any changes on a Radeon 7900 XTX, with programs such as Blender running flawlessly. The project is essentially a drop-in replacement that will process CUDA-related requests correctly so that applications can continue running as if they were on true Nvidia hardware. There are some limitations though, most notably that ZLUDA runs on ROCm 5 instead of 6. ZLUDA can also work on Windows machines.

To use ZLUDA, you can simply grab the latest release files and specify the $LD_LIBRARY_PATH environment variable for the CUDA library files. From there, whatever application you're using should just work. Especially interesting is that Phoronix identified better performance in some instances using ZLUDA versus using the native Radeon HIP in Blender. Janik also found that Geekbench using ZLUDA had better performance in many areas versus using the native OpenCL implementation.

If you want to give ZLUDA a try, you can download it from Janik's GitHub page now. There you'll also find instructions for using it on Windows and Linux, along with known issues and an FAQ. It's not a perfect implementation of CUDA on AMD, but it's currently the only one, and with applications working as well as they do on it, for some people it may already work for exactly what they need.