Getting started

This site follows one desktop with two Arc Pro B60s. If you're building an Intel GPU inference machine, these are the useful starting points.

Set up the machine

The two-B60 setup guide covers the hardware, software versions, startup hangs and configuration used in the October 2026 tests. Use its recorded configuration as a reference; check upstream requirements before using a newer release.

For a current installation, use the Intel XPU section of the upstream vLLM GPU installation guide. It documents a different software version from the build log.

Compare the tradeoffs

The execution-mode comparison measures eager, compiled and graph capture on the same task set. The benchmark table has the other configurations; expand a row's notes for its conditions. Quality scores use different tests across rows.

The measurement guide is a checklist for planning your own comparison.

Follow the work

Lab notes record reviewed milestones. News covers relevant upstream changes and says when they haven't been tested here. The RSS feed includes both alongside original articles.

The notebook JSON feed contains the reviewed result rows behind those entries. Older benchmark-table rows do not yet have public raw logs; they are one machine's observations, not a hardware ranking.

For corrections, send the page URL and the version or result you're questioning to contact@localxpu.com.