b9813

ggml-org/llama.cppb9813Jun 26, 2026by github-actions[bot]same day as b9811

What changed since b9811

medium risk

Adds Intel Xe1 architecture detection and enables coopmat1 on Xe-LPG Plus GPUs in the Vulkan backend.

New features

  • New INTEL_XE1 Vulkan architecture enum for classifying Intel Xe1-class GPUs.
  • Cooperative matrix (coopmat1) acceleration is now enabled on Intel Xe-LPG Plus hardware.

Improvements

  • Improved Vulkan performance for Intel Xe1 GPUs by unlocking coopmat1 paths.
  • Added a Windows driver version check to gate the new Intel coopmat1 path safely.
  • Removed an earlier driver workaround after renaming INTEL_PRE_XE2 to INTEL_XE1.
  • Cleaned up bf16 handling and trailing whitespace in the Vulkan backend code.

Upgrade notes

No configuration or API changes are required; the delta is confined to the Vulkan backend. Intel GPU users on Xe1/Xe-LPG Plus should see faster inference after upgrading, while other backends are unaffected. The prior conv2d coopmat2 compiler workaround from b9811 remains in place.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Adds support for Intel Xe1 architecture and enables cooperative matrix operations (coopmat1) on Intel Xe-LPG Plus GPUs, improving performance for Intel GPU users.

Key Highlights

  • Added INTEL_XE1 architecture enum
  • Enabled coopmat1 on Intel Xe-LPG Plus
  • Improved Intel GPU support

Release Notes (excerpt)

<details open> vulkan: add INTEL_XE1 arch enum and enable coopmat1 on Intel Xe-LPG Plus (#24404) * vulkan: add INTEL_PRE_XE2 arch enum and enable coopmat1 on Intel Xe-LPG Plus (1/3, Xe1-ARLH) Co-authored-by: Xia, Jie <jie.xia@intel.com> Co-authored-by: Liu, Russell <russell.liu@intel.com> * Address comments of bf16 and trailing whitespace * Rename INTEL_PRE_XE2 to INTEL_XE1 and remove driver workaround * Add Windows driver check --------- Co-authored-by: Xia, Jie <jie.xia@intel.com>…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories