Primary documentation reviewed September 10, 2026. Linux host guidance and separate application choices; no new installation or B70 inference test. Versioned upstream support is distinguished from validation on our hardware.
Start with the job
Choose the application before assembling its dependencies. A GGUF chat session, a concurrent inference server, and an image workflow need different validation. These are starting routes to evaluate, not a ranking of speed or quality.
| Your task | Starting route | First useful check |
|---|---|---|
| Run a GGUF model or one conversation | llama.cpp SYCL | Device offload, readable output, then a second request |
| Serve simultaneous requests | Intel LLM Scaler or upstream vLLM XPU | Correct responses at your intended concurrency |
| Run a Python model | Native PyTorch XPU | Required operations execute on the intended device |
| Generate images or video | ComfyUI / Omni guide | A saved output and repeatable workflow on one card |
| Adapt a CUDA-first project | CUDA-to-Intel decision guide | Identify custom kernels before changing dependencies |
For a first inference test, continue with Build a baseline you can trust. Keep the troubleshooting log nearby when a symptom looks familiar.
Choose one host package route
Intel’s Linux installation overview separates a pinned OMIX stack from the Ubuntu PPA route. It explicitly warns against mixing the PPA’s packages into OMIX. Record which route the machine actually uses before adding an application environment.
| Route | Why consider it? | What to preserve |
|---|---|---|
| OMIX | Intel validates a defined combination of components for supported hardware | Exact OMIX release and its host/component matrix |
| Ubuntu PPA | Your application requires a deliberately selected newer component set | Repository origin and exact installed versions |
| An existing working host | Establish a baseline before changing the machine | Current package record and an isolated application environment |
A container isolates much of the application stack; it does not supply a replacement host kernel or remove the need for compatible driver access. A new virtual environment likewise cannot repair an unsupported host driver.
Read the host and container columns separately
The OMIX support matrix, checked September 10, lists B70 validation for OMIX 0.1–0.3. For 0.3, it lists an Ubuntu 26.04 host, with Ubuntu 24.04 or 26.04 containers.
The installation page also names Ubuntu 24.04.4 and 24.04 HWE among host prerequisites. That is broader than the versioned matrix. If evaluating a fresh baseline from these documents, Ubuntu 26.04 is the common documented host choice. Support for a particular 24.04 host needs clarification against the chosen release. This is a documentation comparison, not our tested installation recipe.
Intel’s Xe hardware table identifies B70 as PCI ID E223 using the Xe driver. Its distro/kernel support table is more useful than assuming that any system with a sufficiently large kernel version will work.
Keep application versions together
For llama.cpp, save the source commit or binary build, build flags, compiler/runtime versions, and any local diff. Use the SYCL instructions belonging to that revision. An April or June environment in our benchmark records is historical evidence, not today’s dependency list.
For vLLM, choose either upstream XPU packaging or an Intel LLM Scaler release and follow that release’s instructions. Upstream now documents official XPU images from v0.26.0. Its current wheel path specifies Python 3.12 and an XPU triton compatibility package; package names alone are not a reason to apply an old uninstall workaround. See the XPU installation source.
Intel’s release list names intel/llm-scaler-vllm:0.26.0-b2 as the September release. Save the resolved container digest as well as its tag. A release note is a reason to retest a workload, not evidence that our older benchmark numbers now apply to it.
For creative workflows, distinguish a published Omni image from current source-build instructions. The creative guide explains that boundary.
Make the environment reproducible
Before measuring anything, save a short record with these fields. Unknown values should stay unknown until collected.
Test date:
Host OS / kernel:
GPU model / physical device mapping / available memory:
Host package route:
Compute runtime / Level Zero loader:
Application commit or image tag + digest:
Compiler / framework / dependency versions:
Local changes or custom nodes:
Model revision / weight files / hashes:
Launch configuration / device selection:
Prompt or workflow / seed / actual workload size:
Observed output / errors / raw timing files:For repeated runs, keep the same weights and workload. Separate loading, first-use compilation, and warmed execution. Verify response content before keeping a fast result. The baseline checklist describes what a first useful pass looks like.
Windows and experimental alternatives
Windows owners should start at Intel’s B70 download hub and use the appropriate driver track and board guidance. The Linux package choices above do not apply to Windows.
AI Playground is a graphical option to evaluate; its Linux support is experimental. Ollama documents experimental Vulkan support, a separate route from SYCL. Check the application’s exact backend and model requirements before comparing results.
Older search results may lead to IPEX-LLM, archived January 28, 2026. Treat those installation recipes as historical. That archive status does not describe native PyTorch XPU.