ZLUDA aims to run unmodified CUDA applications on non-NVIDIA GPUs. Its README calls it a drop-in replacement with near-native performance. That is the project’s claim; its development posts describe a narrower set of working applications and continuing compatibility work.
What gets translated
The project includes a compiler that translates NVIDIA PTX into code for AMD GPUs. Its development update explains that compilation happens when an application loads a GPU module. A kernel cache avoids repeating that work for code already compiled on the machine.
Supporting an instruction set is only part of running an application. CUDA programs also call host APIs and performance libraries. The same update describes initial support for selected library operations and acknowledges gaps. An application may load successfully and still reach an unsupported operation later.
This differs from source-level porting with HIPIFY. A port changes application source and recompiles it for the target. ZLUDA tries to supply the interfaces an existing CUDA application expects. Its compatibility depends on which parts of CUDA that application actually uses.
What the project has demonstrated
The Q3 2025 post, which announced ZLUDA 5, says two llm.c tests run when built “without Multi-GPU and without Flash Attention”. Flash Attention was blocked by missing MIOpen APIs. The same post reported that the CUDA backend of llama.cpp ran, with preliminary performance within range of Phoronix’s ROCm results.
The Q4 2025 post, dated 13 January 2026 and linked from the project’s news index, reports “complete llama.cpp support” and says performance is nearly identical to the native ROCm backend. That is the author’s own comparison, not an independent benchmark. The post also says “ROCm 7 works now” and lists “PyTorch support underway”, with vLLM as the primary objective. The project missed its goal of PyTorch support by the end of 2025, the author says there is not much PyTorch support to show yet, and no new major release accompanied the post. A KataGo trace ran thousands of times slower than on NVIDIA GPUs because of a slow MIOpen convolution implementation, which the author says ZLUDA cannot fix.
The Q1 and Q2 2026 post, dated 29 June 2026, announces version 6 and reports fixes driven by traces from PyTorch users. It does not announce complete PyTorch compatibility. The news index lists nothing newer. The GitHub repository does: prerelease v7-preview.13 was published on 2 October 2026, and recent commits add cuDNN and cuFFT functions, the latter through hipFFT. No post describes version 7, so what it changes is undocumented in the sources reviewed here.
Funding and scope
The project’s “third life” post, from October 2024, says the author found a commercial organisation willing to fund development. It describes a rollback to the pre-AMD codebase and a rebuild that prioritised machine-learning workloads and targeted RDNA1 and newer non-server AMD GPUs. Raytracing was dropped and Windows support was reduced.
The rollback has a stated cause. The README revision of 6 August 2024 says the earlier code was taken down at AMD’s request, that it had been released with AMD’s approval by email, and that AMD’s legal department later said the email was not legally binding. The author adds that he received no legal threats or communication from NVIDIA. Wikipedia’s CUDA article, citing Phoronix, says Intel contracted the author in 2021 and AMD in 2022, that neither company released the software officially for lack of a business use case, and that AMD’s contract let him release his AMD code independently.
The Q1 and Q2 2026 post reports that commercial funding ended and development became the author’s weekend project. It announces texture support and preliminary PhysX work, and improves Windows library loading. Those additions show why an earlier roadmap should not be read as a permanent support promise.
For Windows, the author describes a practical dependency problem. The GPU driver does not supply every ROCm performance library an application needs. ZLUDA now reports missing libraries more clearly, but users still have to install them.
Licensing questions
ZLUDA publishes its source, but its source licence alone cannot settle whether a particular use of NVIDIA material complies with NVIDIA’s terms. The CUDA Toolkit EULA restricts specified reverse engineering of the SDK and of SDK-generated output for translation to a non-NVIDIA platform. Applying those clauses requires examining the particular implementation and use.
This summarises published documents. It is not legal advice.
FAQ
Can I run any PyTorch workload through ZLUDA?
No project update promises that. The Q4 2025 post lists PyTorch support as underway, with vLLM first, and the June 2026 post reports fixes from user-submitted traces. Test the particular workload before relying on it.
Which GPUs does ZLUDA target?
The rebuild post targets RDNA1 and newer non-server AMD GPUs. That is the stated scope of that development plan, not a guarantee for every application on those devices.
Is ZLUDA legal to use?
The public project documentation cannot settle that for every use. The answer depends on the applicable licences and the materials the application or translation layer uses.
Is development still active?
The newest news post is from June 2026 and says commercial funding ended and the author develops ZLUDA as a side project with less frequent updates. The repository had commits and a v7 prerelease on 2 October 2026.