One of the unexpected twists this year was after several years of AMD quietly funding the [1]ZLUDA developer for enabling unmodified CUDA applications to run on AMD GPUs at near-native performance, [2]the ZLUDA atop AMD HIP code was made available and open-source following the end of the AMD contract. But then later on [3]that ZLUDA code was taken down at the request of AMD. Back in October ZLUDA then decided to pursue a new life as [4]an open-source multi-GPU CUDA implementation with an emphasis on AI workloads. Now as a New Year's Eve surprise, ZLUDA v4 was released as the first step to that new codebase.

Following AMD requesting the removal of the ZLUDA source code they funded, ZLUDA developer Andrzej Janik has been working on a new rewrite that also wouldn't be limited to just AMD GPUs but could be used across vendors for CUDA on non-NVIDIA GPUs more broadly. ZLUDA v4 released this New Year's Eve as the first release to be made off that new codebase following the AMD rollback.

But before getting too excited, ZLUDA v4 support is quite limited and the only known intended program to be currently working is Geekbench with its CUDA benchmarks. It will take more time before seeing broader CUDA application support atop the new ZLUDA codebase. Janik's announcement for ZLUDA v4 simply reads:

"This is the first release post-rollback and is very limited: only Geekbench is supported"

In any event those interested in ZLUDA v4 for Windows and Linux non-NVIDIA systems can find it via [5]ZLUDA on GitHub .



[1] https://www.phoronix.com/search/ZLUDA

[2] https://www.phoronix.com/review/radeon-cuda-zluda

[3] https://www.phoronix.com/news/AMD-ZLUDA-CUDA-Taken-Down

[4] https://www.phoronix.com/news/ZLUDA-Third-Life

[5] https://github.com/vosen/ZLUDA/releases/tag/v4