ZLUDA Version 6 Now Lets AMD GPUs Run NVIDIA's CUDA Software

The vosen/ZLUDA project released Version 6 on 30 June 2026. ZLUDA is an open-source compatibility layer — essentially a translation tool — that lets software written for NVIDIA GPUs run on AMD GPUs without any modification to the original code.

Here's how it works: most AI and machine learning software is built specifically for NVIDIA's CUDA platform, which is deeply integrated into NVIDIA's ecosystem. When you want to run that same software on an AMD GPU instead, you normally have to rewrite or recompile it. ZLUDA intercepts the calls that software makes to CUDA and translates them on the fly for AMD's ROCm platform. Think of it as a real-time translator sitting between the application and the hardware.

Version 6 targets AMD's ROCm 6.2, which is AMD's current GPU software stack. This is the 184th release of the project, published under an MIT license, meaning anyone can use it, modify it, or include it in commercial products without restrictions.

For the last decade, running GPU-accelerated workloads — machine learning inference, scientific simulations, rendering — across different GPU vendors has been a practical headache. CUDA became the de facto standard, so most AI frameworks and tools now assume NVIDIA hardware, even though the underlying mathematics works on any GPU. ZLUDA removes that assumption: AMD users can now run a much wider range of unmodified CUDA software directly on their hardware.

The practical upshot is real. AMD's ROCm has matured and now powers serious data-center GPU deployments, yet most ML tooling still expects CUDA out of the box. When a compatibility layer keeps pace with current ROCm releases — version 6.2, not a years-old version — AMD users gain access to substantially more working software without modification.

There are real limits worth understanding. CUDA is not a single simple interface; it covers compute kernels, memory management, graph APIs, and various extensions. No translation layer covers all of it perfectly, and the performance cost of translation varies depending on the workload. If you are running latency-sensitive AI inference or workloads that demand high memory bandwidth, you should test your specific software on AMD hardware rather than assume it will run at the same speed as on NVIDIA. The project's documentation is the right place to understand what is currently supported and what is not.

The broader picture is a GPU market where NVIDIA's hold is strong but loosening. HIP, OpenCL, SYCL, and tools like ZLUDA all approach the same problem: reducing the friction of moving compute workloads between GPUs or vendors. Each has trade-offs. Native ports to HIP give you the best performance but require code changes. ZLUDA prioritises keeping your code unchanged, accepting some translation overhead as the cost. For teams deciding whether to use multiple GPU vendors in their infrastructure, a well-maintained, MIT-licensed compatibility layer that tracks current AMD releases is a practical option to consider.

Version 6 is available now from the vosen/ZLUDA GitHub repository.