Reference

GPU compatibility and testing

Review GPU development testing across Windows and Linux, then check the runtime included in your exact package.

GPU testing completed since RC1

Pūnaha development builds have completed functional and endurance testing across NVIDIA CUDA, AMD HIP/ROCm, Intel SYCL and Vulkan. Results below distinguish Windows from native Ubuntu and retain the incomplete Ubuntu Vulkan endurance result. Evidence was reviewed on 12 November 2026.

Recorded GPU development testing
Acceleration technologyWindowsNative Linux
NVIDIA CUDAWindows Server 2025: Functional and endurance testing passedUbuntu 22.04: Functional and endurance testing passed
AMD HIP/ROCmWindows Server 2025: Functional and endurance testing passedUbuntu 24.04: Functional and endurance testing passed
Intel SYCLWindows 11: Functional and endurance testing passedUbuntu 24.04: Functional and endurance testing passed
VulkanWindows 11: Functional and endurance testing passedUbuntu 24.04: Functional testing passed; endurance testing incomplete

The Ubuntu Vulkan endurance run stopped following an observation-check failure. It is not recorded as a completed endurance pass or evidence of a GPU crash. The Windows Intel development review also retains unresolved memory-observation findings alongside its functional and endurance passes.

What these results establish

These are recorded development tests on the operating systems shown. NVIDIA and AMD results come from component tests. Windows Intel results include an installed development Product campaign. They do not qualify every device or driver, another Linux distribution, or the final 1.0.0 installer. A passed development test and a runtime included in a released package are separate facts.

Runtime supplied with the 1.0.0 installers

The inspected Ubuntu 24.04, Debian 13, RHEL 9, SLES 15 SP7 and generic Linux installers contain the CPU inference and embedding runtime. They do not contain GPU workers. Their CPU-only inventory does not erase the development GPU results above. A graphics driver alone cannot supply the missing packaged runtime.

The Windows MSI has been published with the CPU inference runtime. Check the exact installed package before claiming a GPU worker is available. See platform and installer testing and the 1.0.0 package list.

Choose processing on an installation

  1. Identify the exact package, operating system and bundled runtime.
  2. Confirm that the server has a compatible physical GPU and driver when a GPU runtime is supplied.
  3. Review the processing selection for the model, workload and server. Generation and embeddings have separate selections.
  4. Test the intended operation and inspect the actual backend used. Loading a model into memory alone does not prove successful execution.

Windows HIP remains an experimental, manual opt-in backend in the reviewed source. Automatic selection excludes it. GPU-only processing must not silently change to CPU; a GPU-first policy may use CPU fallback where permitted.

Continue with your model

Follow the first model-to-chat tutorial, the knowledge-chat tutorial and worker recovery. Their written instructions target 1.0.0. Older recordings retain their RC1 identity and do not replace the later testing results here.

Pūnaha Docs

Search the guides

Enter at least two characters.

    Product screen

    View the full screenshot