ZLUDA 6: Running CUDA on AMD GPUs Is Now a Hobby Project

ZLUDA 6 released today with a surprising set of features: 32-bit PhysX support, texture implementation for Blender compatibility, and improved Windows tooling. The project lets you run unmodified CUDA applications on AMD GPUs, and with this release, the development direction has fundamentally changed.

Last updated: June 30, 2026

Official Sources#

What Is ZLUDA?#

ZLUDA (the name is Polish for "mirage" or "illusion") translates CUDA calls to run on non-NVIDIA hardware. You take a CUDA application, run it through ZLUDA, and it works on AMD GPUs via ROCm.

The project has had a complicated history with AMD. The developer was funded by AMD for several years to help break the CUDA moat for ML workloads. That arrangement ended, and AMD's legal team actually went after the developer for releasing that code as open source. The current version is rebuilt from a pre-AMD-funded codebase.

What's New in Version 6#

From the release post:

PhysX Support (Pre-alpha): 32-bit PhysX now works on AMD GPUs. This means older games with PhysX effects - debris, flames, particle systems - can render those effects on AMD hardware. Fluid simulations are still glitchy, but basic effects work.

Texture Support: Basic texture implementation enables Blender compatibility. This was previously impossible because CUDA textures weren't translated.

Windows Improvements: Better error messaging, automatic performance library loading, and more user-friendly configuration. Windows still requires manual ROCm installation since AMD doesn't bundle it with drivers.

ML Enhancements: Multiple compiler fixes and new GPU instructions supporting PyTorch workloads.

The Hobby Project Pivot#

The most interesting part of the release is the project's new direction. From the author:

"ZLUDA development is no longer commercially funded, so it's back to being my weekend project. This means that the priority is no longer what makes commercial sense, but what I find the most entertaining. That's why the sudden addition of textures, PhysX and better Windows support."

This explains why a project previously focused on ML inference suddenly added gaming features. PhysX support doesn't make commercial sense - it helps with old games on AMD GPUs - but it's a fun technical challenge.

The flip side is that updates will be less frequent and less predictable.

What HN Is Saying#

The Hacker News discussion was smaller (133 points, 12 comments) but touched on some interesting points.

Legal questions: One commenter asked if ZLUDA violates NVIDIA's license terms. The answer is unclear. NVIDIA's EULA may prohibit running CUDA on non-NVIDIA hardware, but whether that's actually legally enforceable depends on jurisdiction, how ZLUDA was built, and what exactly NVIDIA would go after. The project has survived this long without legal action.

LLM use cases: Someone asked how ZLUDA compares to Vulkan for running LLMs on AMD hardware. The consensus: Vulkan and OpenCL paths have matured significantly. Frameworks like Unsloth have made the CUDA moat less relevant for ML specifically, since you can often get native AMD support without a translation layer.

The NVIDIA irony: The discussion noted that NVIDIA briefly considered dropping 32-bit PhysX support on their own 5000 series cards. They reversed course after backlash, but there was a period where people wondered if AMD users with ZLUDA would have better PhysX support than NVIDIA users with new hardware.

Z-L-U-D-A pronunciation: The name is Polish. "Zluda" means mirage or illusion, and CUDA is Polish for "miracles." Layers of meaning.

Practical Considerations#

If you're considering ZLUDA for real workloads:

For ML/LLMs: The native ecosystem has caught up. Frameworks increasingly support AMD directly via ROCm. ZLUDA remains useful for CUDA-only libraries that haven't been ported, but check if native support exists first.

For gaming with PhysX: The new support is pre-alpha. Expect glitches, especially with fluid simulations. But if you have old games that are unplayable on AMD due to missing PhysX effects, this could help.

For Blender: Texture support enables CUDA rendering paths on AMD GPUs. Worth testing if you've been stuck on OpenCL or CPU rendering.

For Windows: You need to manually install ROCm. AMD doesn't ship it with consumer drivers. This adds friction compared to NVIDIA's just-works CUDA distribution.

For production: The project is unfunded and updated on the author's entertainment schedule. Factor that into reliability calculations.

My Take#

ZLUDA is a fascinating technical project that exists because NVIDIA's CUDA moat created artificial lock-in. The fact that a single developer can translate CUDA to ROCm shows the moat was always more about ecosystem and inertia than fundamental technical barriers.

The shift to hobby-mode is both limiting and liberating. We probably won't see enterprise-grade support or rapid bug fixes. But we're getting features that "make commercial sense" would never prioritize - like making old PhysX games work on AMD hardware.

For developers, the bigger picture is that the CUDA moat is eroding from multiple directions. Native framework support, translation layers like ZLUDA, and Apple's work on MLX all chip away at NVIDIA's lock-in. The question isn't whether alternatives will exist, but which will mature fastest for your specific workload.

ZLUDA 6 is a good release for what it is: a passionate side project that solves real problems for people stuck with AMD hardware and CUDA-only software. Just don't build production infrastructure on it.

FAQ#

Does ZLUDA work with all CUDA applications?#

No. Coverage is partial and depends on which CUDA APIs the application uses. Basic compute works well. Textures are new and basic. Some features remain unimplemented.

Is ZLUDA legal?#

The answer is unclear. NVIDIA's EULA may contain restrictions, but enforceability varies by jurisdiction and depends on how ZLUDA was built. The project has operated for years without legal action.

Should I use ZLUDA for ML workloads?#

Check if native AMD support exists first. Many frameworks now support ROCm directly, which is more reliable than translation. ZLUDA is most useful for CUDA-only libraries without native AMD ports.

Will ZLUDA support Intel GPUs?#

The post doesn't mention Intel. Current focus is AMD GPUs via ROCm. Intel support would require different backend work.

Continue Reading#

• AMD MI355X vs NVIDIA B200 vs B300 for Open-Weight Serving in 2026
• Distilling an LLM on One GPU: Offline Top-K Logits and a Fused Chunked KL Loss
• Gleam Moves to Tangled: What the ATProto Code Forge Means for Developers

Sources#

• ZLUDA Q1/Q2 2026 Update
• HN discussion (48730713)
• Tom's Hardware coverage