GPU compatibility and supported runtimes

Check your RC1 package first

This guide covers Punaha 0.1.0-rc.1, distribution rebuild-20261012, reviewed on 14 October 2026. Use the matching downloads and release validation notes.

No GPU model is yet recorded as having completed GPU acceptance for this published rebuild. Windows includes GPU runtimes. Hardware availability and earlier test results do not establish acceptance of these new binaries.

Runtimes included with RC1

A runtime backend is the component that runs your model on a processor. These packages use the reviewed llama.cpp b10212 runtime.

Package	Included backends	GPU position
Windows x64	CPU, Vulkan and Intel SYCL	GPU processing is available for compatible hardware and drivers, subject to successful model tests.
Ubuntu amd64, Debian 13 amd64 and RHEL 9 x86_64	CPU	The published Linux packages do not include a GPU backend.
macOS and Apple GPUs	No published RC1 package	Metal support in upstream software is not a Punaha RC1 package.

CUDA and ROCm backends are not bundled. NVIDIA and AMD devices may instead be candidates for the Windows Vulkan backend. Each device, driver and model combination needs testing.

GPU families and their evidence

Earlier test means a recorded Punaha check on an earlier build. Candidate means a possible runtime option that still needs Punaha testing on the exact hardware.

GPU or family	Windows runtime option	Evidence status
Intel Iris Xe Graphics, device 46A6, driver 32.0.101.7092	Vulkan	Earlier test on one Windows x64 machine. Current rebuild acceptance is outstanding.
Intel Arc A770, A750 and A730M, and Arc B580	Intel SYCL	Candidate. These examples appear in upstream hardware guidance, not Punaha acceptance results.
Integrated Intel Arc and Intel graphics in 11th generation Core processors or newer	Intel SYCL, subject to device and driver compatibility	Candidate. A processor generation alone does not establish readiness.
Intel Data Center Max and Flex families	Intel SYCL	Candidate from upstream guidance. No Punaha acceptance result is recorded for these families.
NVIDIA GPUs with a compatible Vulkan driver	Vulkan	Candidate. No specific NVIDIA model is verified for this rebuild.
AMD GPUs with a compatible Vulkan driver	Vulkan	Candidate. No specific AMD model is verified for this rebuild.

The Intel candidate list comes from llama.cpp b10212 SYCL hardware guidance. It describes upstream capabilities, not the exact composition or acceptance of Punaha packages.

Upstream lists other backends and operating systems. Follow this page's package table when using RC1. An unlisted GPU is unverified rather than automatically proven incompatible.

What the earlier Iris Xe test proves

On 9 October 2026, an earlier RC1 test build passed nine Gemma generation checks across three fresh Vulkan worker starts on the Iris Xe configuration above.

The checks required GPU execution without CPU fallback. They establish that specific test result, not every Iris Xe card, driver, model, embedding workload or sustained production workload.

The published rebuild includes an automatic compatibility rule for this device, driver and runtime combination when exactly one adapter is present. Other drivers and multiple adapter configurations are outside that rule.

Check the compatibility notice in platform Configuration. Do not copy internal runtime settings to another machine. This rule still requires successful tests with the current package.

Drivers, paths and memory

• Install the appropriate GPU driver from your hardware vendor. The Product runtime does not replace the operating system's graphics driver.
• For Intel SYCL on Windows, use an ASCII only Local runtime path. The default installation path uses Program Files\Punaha.
• Use the runtime supplied with the matching Product package. Runtime updates require a reviewed Product package. Downloading model weights is a separate step.
• Check both system RAM and GPU memory. Model size, quantisation, context length and concurrent work affect capacity. There is no single memory requirement covering every model.
• Integrated graphics share system memory. A recognised graphics card or a loaded model does not prove that sufficient memory remains for your workload.

Verify your own GPU

1. On Windows, open Device Manager and expand Display adapters. Open the GPU's Properties and record its name and driver version.
2. When comparing the Iris Xe entry, use Details and Hardware Ids to check the device code. Keep full machine identifiers private.
3. Confirm the RC1 distribution, selected server, runtime and model file. Record the model's quantisation and the workload you intend to run.
4. Open AI → Models. Check Backend readiness separately from In memory.
5. Open Operations → Model testing and run a manual test for the intended model, server and workloads when capacity is available.
6. Test generation and embeddings separately if you need both. Inspect the actual backend and any CPU fallback. A successful CPU answer does not prove GPU execution.
7. Run a small real chat or knowledge workflow with safe input. Check answer quality and repeated operation before increasing the workload.

A manual GPU only policy prevents silent CPU fallback, but it can wait or fail when GPU capacity or health is unsuitable. Choose the processor policy for your workload.

Repeat relevant tests after changing the driver, runtime, model or hardware. Retain the package revision and results with your installation records.

If a GPU is unavailable

Read Saved backend health in platform Configuration. Follow the GPU recovery guide for held backends and manual retesting.

Use CPU processing where your policy permits it while investigating. Restarting repeatedly or deleting saved health records does not establish GPU readiness.

Next step

Follow Run local AI models, then try your first Gemma chat. Keep the package's validation limits with your own hardware test results.