Thanks to ZLUDA, you can run Nvidia CUDA applications on AMD Radeon GPUs. The underlying project was started by AMD but completed as an open source tool.

Nvidia is dominant in the AI ​​space not just because of its powerful hardware. The associated ecosystem surrounding Nvidia CUDA contributes to its popularity. CUDA is good, but it’s also from Nvidia, which has no interest in making it an open system. Intel CEO Pat Gelsinger recently announced that the entire industry is actually against Cuda.

Hating CUDA is one option, embracing CUDA is another. Developer Andrej Janik started a project at Intel years ago to make Intel GPUs compatible with CUDA applications. Janik later moved from Intel to AMD to continue the project as part of the ROCm platform. ROCm is AMD’s alternative to CUDA, but the platform is currently significantly less popular.

Open source

Ultimately, AMD also stopped the plans, but Janik was allowed to continue working. He expanded his previous work to ZLUDA. ZLUDA works within ROCm and ensures that CUDA applications continue to run on AMD hardware. ZLUDA acts as a software layer that translates the instructions. Depending on the application, this translation provides a measurable performance improvement. For example, Blender renders 12 percent faster via CUDA and ZLUDA on AMD than the standard Radeon HIP platform.

In order to make ZLUDA practically usable, maintenance work is required. Janik won’t be able to do this alone. The software’s open source license enables help and investment. It is not entirely clear why AMD itself stopped using ZLUDA. The manufacturer probably does not want to position its ROCm platform as a translation layer for the competing platform, but as a worthy and independent alternative. If you want to get started yourself, you can find ZLUDA on GitHub.