ZLUDA Version 5-preview.43 was released today as this open-source CUDA implementation for use on non-NVIDIA GPUs, with one of the current focuses being on enabling CUDA on AMD Radeon GPUs with ROCm.
Last week was a status report for the current incarnation of ZLUDA for bringing CUDA compatibility to non-NVIDIA GPUs. Good progress continues to be made and there are now two full-time developers working on the effort.
ZLUDA lead developer Andrzej Janik today released ZLUDA 5-preview.43. This new ZLUDA 5 build now enables link-time optimizations (LTO) for better performance, improves the project’s build system, implements mode tracking for AMD GPUs, fixes ROCm 6.4 compatibility issues, implements additional CUDA APIs, starts working on NVIDIA PhysX 32-bit support, adds support for multiple return arguments, and brings a variety of other features.
This is the first new ZLUDA preview release since the end of 2024 and thus there is a lot in tow. Among the new CUDA additions are for cuGetProcAddress and cuGetProcAddress_v2, runtime_callback_hooks_fn2, cuModuleGetLoadingMode, cudart_interface_fn2, and cuLibraryLoadData. There is also additional CUDA driver APIs in place for allowing simple CUDA samples to run.
For now this ability to run unmodified CUDA apps on non-NVIDIA GPUs is focused on AMD GPUs of the Radeon RX 5000 series and newer. Besides CUDA code samples, GeekBench has been one of the early targets for testing.
Downloads and more details on this new ZLUDA 5 preview build for Linux and Windows systems via ZLUDA on GitHub.
