Following this morning’s Intel Graphics Compiler 2.30.1 release, the Intel Compute Runtime 26.09.37435.1 is now available for this open-source OpenCL and Level Zero compute stack for Intel graphics hardware on Windows and Linux.
Intel Compute Runtime 26.09.37435.1 brings a variety of feature work for Level Zero, continued enablement around Xe3P Nova Lake and Crescent Island platforms, and various other new features and performance optimizations.
Intel Compute Runtime public releases still lack any concise release highlights / feature overview, but in digging through all the commits making up the Intel Compute Runtime 26.09.37435.1 release, some of the highlights include:
– Working on a new “modules package binary” feature for Level Zero. I don’t think this feature has been formally announced yet and the patches still seem to be ongoing, but likely something to look out for in the future with this modules packages functionality.
– Enabling multi-device Unified Shared Memory (USM) pool support for Level Zero.
– 32-bit stateless support to complement the existing 64-bit stateless mode now called wide stateless.
– VM bind handling support that now works with the upstream Linux kernel’s Xe driver.
– Various other memory pool enhancements.
– The defer backing feature is now enabled by default.
– Tuning and other work for the upcoming Crescent Island hardware.
– Enabling Ultra Low Latency Scheduling (ULLS) on BCS for Nova Lake S platforms.
– Enabling Unified Shared Memory (USM) allocation recycle for Xe HPC to help with performance.
– The USM host manager is now added to the OpenCL driver path.
– A fix for L1 cache flushing on Xe3P platforms.
Downloads and more details on the new Intel Compute Runtime release via GitHub. I will be trying out the new IGC and Intel Compute Runtime shortly in looking to see if there are any measurable performance gains for Battlemage or Panther Lake at least.
