# Intel Compute Runtime 26.40.40196.17 fixes local memory waste on Xe/Arc GPUs

> **Key takeaway:** The release eliminates unnecessary per-process local memory allocation for mid-thread preemption surfaces on unused Intel GPU devices by unconditionally deferring their backing regardless of memory pressure.

**Published:** 2026-11-09T08:42:01+00:00 
**Source:** Intel Graphics Compute Runtime (Xe / Arc) 
**Category:** kernel-os 
**Canonical URL:** https://fosswire.org/news/intel-compute-runtime-26404019617-fixes-local-memory-waste-o.html 

## Summary
Intel has shipped Compute Runtime version 26.40.40196.17, a targeted fix for memory efficiency on Xe and Arc discrete GPUs. The change addresses how mid-thread preemption surfaces are backed in local device memory. Previously, these surfaces received immediate physical backing whenever system memory pressure was low, causing every compute context to consume local memory on every available GPU device at initialization time, even for devices the process would never actually use.

The fix ensures that when deferred backing is enabled, preemption surfaces are always deferred without consulting the memory pressure threshold. This prevents wasted allocations since the hardware only writes to these surfaces when thread preemption actually occurs. The change is tracked internally as GSD-13830 and authored by Intel engineer Neil R. Spruit.

## Memory allocation behavior and preemption surface changes
- **Deferred backing logic:** The memory pressure check previously granted immediate backing to allocations when local memory usage was low; this threshold check is now bypassed specifically for mid-thread preemption surfaces.
- **Preemption surface lifecycle:** These surfaces are created for every compute context during device initialization, but hardware only writes to them when a thread is actually preempted, making eager backing wasteful.
- **Multi-device waste:** Each process allocated local memory for preemption surfaces on every GPU device in the system, including devices it never submitted work to, multiplying memory overhead in multi-GPU configurations.
- **Scope of fix:** The change applies only when the existing defer-backing feature is enabled, narrowing the behavioral change to systems already configured for deferred allocation.

## Why it matters
For Linux workstations and servers running Intel Arc or Xe-based discrete GPUs, this reduces per-process local memory overhead and improves multi-GPU scalability, particularly in containerized or multi-tenant compute environments where many contexts may initialize but not utilize all devices. Distribution packagers of Intel's compute stack should note this as a recommended update for users reporting unexplained local memory pressure on systems with multiple Intel GPUs.