Unofficial Intel XPU Community

Intel GPU notes for people trying to run local AI without losing a weekend to driver/runtime weirdness.

This is an unofficial community project. It is not owned by Intel.

Start Here

• Ask for help: https://github.com/steveseguin/Unofficial-Intel-XPU-Community/discussions
• Read the wiki: https://github.com/steveseguin/Unofficial-Intel-XPU-Community/wiki
• Share a benchmark: template
• Run a system snapshot: Linux / Windows
• Research lab repo: https://github.com/steveseguin/b70-optimization-lab

What This Community Is For

Intel Arc and XPU local AI is promising, but setup is still scattered across drivers, oneAPI, OpenVINO, PyTorch XPU, Level Zero, SYCL, vLLM, llama.cpp, Docker, and random forum posts.

This repo is meant to be the stable hub:

• setup guides for Linux and Windows
• Docker notes people can actually run
• benchmark templates and comparable results
• patch notes for vLLM, llama.cpp, OpenVINO, oneAPI, and SYCL
• troubleshooting for drivers, PCIe topology, XPU visibility, and runtime mismatches
• discussion categories for setup help, benchmarks, guides, patches, build photos, and research leads

Quick Links

I want to...	Go here
Ask a setup question	Discussions
Read community notes	Wiki
Get started	Getting Started
Check drivers/runtimes	Drivers and Runtimes
Diagnose XPU visibility	Diagnostics
Share benchmark results	Benchmarks and Results
Try containers	Docker and Containers
Understand 2x B70 use cases	Two B70 Use Cases
See research status	Current Research Snapshot
Deploy MiniMax on 4x B70	Ubuntu 24 recipe

Good First Post

If you need help, open a Discussion and include:

• GPU model and count
• Windows or Linux version
• what you are trying to run
• what guide or command you followed
• what failed
• logs or screenshots if you have them

If you are on Linux, attach output from:

bash scripts/collect_xpu_snapshot.sh

If you are on Windows PowerShell, attach output from:

.\scripts\collect_xpu_snapshot.ps1

What Counts As A Useful Benchmark

A useful result says more than "I got 80 tok/s."

Include:

• model and quantization
• GPU model/count
• engine: vLLM, llama.cpp, OpenVINO, PyTorch, etc.
• prompt length, output length, context length
• batch size or concurrency
• output tok/s and total tok/s if available
• whether quality was checked
• exact command or linked recipe

Use the benchmark template.

Current Focus

Near-term community priorities:

• make B70/B-series setup reproducible
• document Windows and Linux paths clearly
• get Docker/container recipes that work outside one private machine
• collect comparable vLLM, llama.cpp, OpenVINO, and PyTorch XPU results
• track driver/runtime mismatches and fixes
• promote stable recipes from the fast-moving B70 optimization lab

Current Working Recipe

The most complete current community reference is MiniMax M2.7 INT4 AutoRound on Ubuntu 24 with 4x Intel Arc Pro B70s:

• OpenAI-compatible vLLM server on 0.0.0.0:8000
• default served context: 32768 tokens
• short warm decode observed through the API: about 84.1 output tok/s
• prompt/prefill observed through the API: about 1.7k-1.8k prompt tok/s
• near-full request tested: 32408 prompt tokens plus 64 generated tokens
• display/VRAM note: moving display duty to the onboard ASPEED VGA adapter and booting with xe.disable_display=1 reclaimed enough B70 VRAM to promote the served context from 24k to 32k
• current host note: PCIe4 x16 allreduce bandwidth measured 13.79 GB/s; an older faster PCIe5-class reference measured 27.88 GB/s, almost exactly 2x, which plausibly explains much of the 83 versus 89-93 decode gap
• recipe: https://github.com/steveseguin/b70-optimization-lab/tree/main/repro/minimax-m27-b70-110tps-ubuntu24-20260523
• detailed PCIe/prefill note: https://github.com/steveseguin/b70-optimization-lab/blob/main/notes/2026-05-23-current-host-pcie4-prefill-check.md

Treat this as a working starting point, not the final speed ceiling.

Repo Map

• COMMUNITY.md: what belongs here and how to write useful posts
• CONTRIBUTING.md: contribution rules
• community/: suggested discussion categories and welcome text
• docs/wiki/: wiki-ready guide drafts
• docker/: container notes
• scripts/: snapshot/diagnostic helpers
• templates/: benchmark/report templates
• schemas/: structured result schema

Important Caveat

This is community research. Some notes will be experimental, incomplete, or wrong for your exact machine. Prefer posts with commands, versions, logs, and reproducible steps.

Do not post API keys, private model files, proprietary data, or copied vendor docs. Link to sources instead.