ZLUDA allows unmodified CUDA code to run on AMD Radeon GPUs
CUDA code can now run natively on AMD Radeon graphics cards – Thanks ZLUDA
For the past two years, AMD has been secretly funding a project called ZLUDA, adapting the once Intel-backed project to get CUDA running natively on Radeon graphics cards. Strangely, AMD has now discontinued its support for the project, but has allowed it to be shared by its creator as open-source software. This project allows native CUDA code to run on AMD GPUs, and in many cases it it runs faster than AMD’s Radeon HIP code.
ZLUDA started as an effort to get CUDA code running on through Intel’s OneAPI. This project was discontinued and then funded by AMD to achieve the same goal on Radeon hardware. While ZLUDA cannot run Nvidia Optix code on Radeon graphics hardware, it can run CUDA binaries without any changes. This means that many native CUDA applications can now run on AMD hardware.
ZLUDA lets you run unmodified CUDA applications with near-native performance on
ZLUDA is currently alpha quality, but it has been confirmed to work with a variety of native CUDA applications: Geekbench, 3DF Zephyr, Blender, Reality Capture, LAMMPS, NAMD, waifu2x, OpenFOAM, Arnold (proof of concept) and more.
(Data from Pheronix)
ZLUDA Benchmark Performance – CUDA on AMD
Above we can see that ZLUDA allows Radeon graphics cards to run native CUDA code in Blender 4.0. Note that this allows Radeon GPUs to run faster than AMD’s own Radeon HIP code. Below, we can also see how ZLUDA can allow CUDA code to run faster than OpenCL code on AMD GPUs. That said, some applications run slower under ZLUDA.
Moving forward, Andrez Janik, the developer behind the project, plans to continue working on ZLUDA. That said, without financial backing, Janik will only work on improving workloads that he is personally interested in. This includes getting DLSS working on Radeon GPUs.
(Data from ZLUDA’s Github page)
Why has ZLUDA been abandoned?
At this time it is unknown why AMD has stopped funding ZLUDA. The company funded the project for two years, and it is clear that within those two years AMD HIP adoption has increased massively. There is now growing interest in non-Nvidia compute hardware. That has resulted in strong sales for AMD’s latest datacenter GPUs. While native CUDA support on Radeon hardware would be a great feature for AMD, they do not need it to secure sales for their hardware.
You can join the discussion on ZLUDA allowing unmodified CUDA applications to run on AMD Radeon GPUs on the OC3D Forums.