One unexpected development this year is that AMD has been secretly funding ZLUDA developers for several years to enable unmodified CUDA applications to run at near-native performance on AMD GPUs. Later, ZLUDA on AMD HIP code became available and became open source after development was completed. Contract with AMD. However, that ZLUDA code was later removed at AMD’s request. Then in October, ZLUDA decided to pursue a new life as an open source multi-GPU CUDA implementation focused on AI workloads. Now, as a New Year’s Eve surprise, ZLUDA v4 has been released as the first step towards a new codebase.
Following AMD-funded requests for the removal of ZLUDA source code, ZLUDA developer Andrzej Janik has announced a new rewrite that will allow for broader use of CUDA on non-NVIDIA GPUs across vendors, and not just on AMD GPUs. We are working on this. . ZLUDA v4 was released on New Year’s Eve of this year as the first release built from the new codebase after the AMD rollback.
But before you get too excited, support for ZLUDA v4 is very limited and the only intended program currently known to work is Geekbench with CUDA benchmarks. It will take more time for broader CUDA applications to be supported on the new ZLUDA codebase. Here is Janik’s announcement regarding ZLUDA v4:
“This is the first release after the rollback and is very limited. Only Geekbench is supported.”
In any case, anyone interested in ZLUDA v4 for Windows and Linux non-NVIDIA systems can find it via ZLUDA on GitHub.
