2523.78

The 2523.78 release supports the following operating systems:

  • Red Hat Enterprise Linux (RHEL): 8.10, 9.4, 9.6, 9.7, 10.0, and 10.1

  • Ubuntu 22.04 and 24.04

  • SUSE Linux Enterprise (SLES): 15 SP4, 15 SP5, 15 SP6, and 15 SP7

Note

When using Intel® Extension for TensorFlow with this release, ensure that you are using Intel® oneAPI Base Toolkit version 2025.2.144 or earlier, as later versions are incompatible.

To install packages for this release, follow the Installing LTS 2523.x guide for your specific operating system.

Improvements

General

Incorporated the latest security updates to address recent vulnerabilities, enhance protection, and ensure greater system reliability.

Intel® Graphics Driver Backports for Linux* OS (i915)

  • Removed fast retirement of RPS boosting requests to fix a performance regression caused by unintentionally setting dma-latency to 0, which forced all CPU cores to busy-spin on idle and diverted power and thermal budget from application workloads.

  • Rescheduled TD_ATT scanning until after pagefault handling clears the in-pagefault state to avoid leaving attention bits uncleared and contexts unterminated.

  • Fixed a CONFIG_TRANSPARENT_HUGEPAGE issue by using PMD_SIZE instead of a THP.

  • Incremented the EU_ATTN PMU counter for attentions raised during pagefault handling.

  • Deferred global GT resets for ATT events while any engine is inside pagefault handling.

  • Restored the null context guard for fences that do not have hardware contexts.

  • Removed the unused dev variable from i915_sysfs_store.

  • Fixed sysfs GT error attribute callback signatures to avoid store callback mismatches.

  • Updated RAS HBM error information handling by expanding field sizes, improving field layout readability, and allowing administrators to clear the sysfs attributes.

  • Reduced metadata size by stopping page-boundary padding.

  • Tracked the originating engine for each VMA in multi-process captures so captured data is associated with the correct engine.

  • Canceled pending vm_bind work during closure to avoid indefinite stalls when debugging blocks the operation.

  • Started releasing the vm->client reference immediately after its final use for the debugger destroy event.

  • Prevented new VMAs from being inserted into a closed ppGTT.

  • Fixed a deadlock race between VM_UNBIND and eviction by reordering the locking.

  • Started keeping the VM open while performing VM_BIND to avoid binds racing with VM destruction.

  • Started waiting for asynchronous vm_bind completion before forcing VM closure.

  • Added an early check for unevictable resident objects to avoid waiting on objects that cannot be evicted.

  • Prevented execution from reviving a VM that is already closing.

  • Moved VM reference ownership to vm->open so open VMs retain the required reference.

  • Propagated context construction parameters to the backend so settings such as context isolation are preserved.

  • Reworked global reprioritization to process dependent requests in topological order and avoid submission-order inversions.

  • Fixed the execbuf unwind path to avoid leaking clients and freeing active contexts on transient errors.

  • Fixed CPU page-table leaks caused by replacing split PMDs with huge PTE leaves.

  • Filtered coredump active requests by using CURRENT_LRCA so pagefault captures identify the correct active context.

  • Added backport support for kernel version 6.18.

Intel® Graphics Compiler for OpenCL™

  • Added safe handling for addrspacecast of null pointers in ConstantEncoder.

  • Updated handling of Standard Portable Intermediate Representation of Vulkan (SPIR-V) HostAccessINTEL decoration IDs to support both legacy and current values.

  • Renamed d8c32 and d16c32 visaasm types to d8u32 and d16u32 to align with the specification.

  • Fixed transformed intel_sub_group_2d_block_read built-ins to propagate valid tile-width information.

  • Reduced Software Scoreboard (SWSB) compilation time for subroutines by removing unnecessary live-out dependencies for call basic blocks.

Intel® Graphics Compute Runtime for oneAPI Level Zero and OpenCL™ Driver

  • Enabled error capture file debugging in online debug mode, correctly returned offline dump context IDs, and allocated SIP kernels per context ID.

  • Disallowed online debugging on Intel Data Center GPU Max Series when more than one Compute Command Streamer (CCS) is selected.

  • Fixed a leaked PRIME file descriptor in createGraphicsAllocationFromExistingStorage.

  • Validated ZEX_NUMBER_OF_CCS values and made zeInit fail with ZE_RESULT_ERROR_UNINITIALIZED when the value is invalid.

  • Extended parseUSMType to support additional return types.

  • Corrected split Blitter Command Streamer (BCS) command estimation.

  • Added custom context IDs for offline dump capture and set OsContext for SBA tracking buffer allocation.

  • Used consumeFd to avoid double-closing PRIME file descriptors.

  • Removed RTLD_DEEPBIND from dlopen calls.

  • Corrected the enum value for OpenCL function tracing types.

  • Ensured command buffers are not marked read-only when debugging is enabled, even if the debugger is not yet initialized.

  • Ignored unknown ze information attributes.