Release Notes: Rolling

This document outlines changes introduced to the Intel® software for general-purpose GPU capabilities in rolling releases. As the software includes several different projects, the changes for each release are grouped by project.

Support for each rolling release continues only until the next rolling release becomes available, with no updates provided for previous rolling releases. Therefore, we recommend upgrading to the latest rolling release as soon as it becomes available. To install packages for the latest rolling release, refer to the installation guide for your distribution. For a list of packages published on repositories.intel.com/gpu for each release and operating system, see Provided Packages.

2025-07-24

The 2523.12 release supports the following operating systems:

Red Hat Enterprise Linux (RHEL): 8.10, 9.4, and 9.6
Ubuntu 22.04 and 24.04
SUSE Linux Enterprise (SLES): 15 SP4, 15 SP5, and 15 SP6

Known issues

In this release, the OpenCL compiler enforces stricter type conversion checks. Kernel code that implicitly converts a global pointer to an integer without an explicit cast may now fail to compile with an “incompatible pointer to integer conversion” error. Most applications are unaffected, but if you encounter this error, update your kernel code to add an explicit cast or compile with the -Wno-error=int-conversion flag.

Features

Intel® XPU Manager and XPU System Management Interface 

Added support for updating specific GPU firmware in recovery mode.
Introduced security improvements.

Improvements

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

Improved GPU error reporting by including UUID resources for better diagnostics.
Enhanced responsiveness during memory management tasks.
Refined tbb thread handling to improve scheduling efficiency, avoid redundant parking during cancellations, and ensure proper wake-up behavior.

2025-07-02

The 2523.10 release supports the following operating systems:

Red Hat Enterprise Linux (RHEL): 8.10, 9.4, 9.5, and 9.6
Ubuntu 22.04 and 24.04
SUSE Linux Enterprise (SLES): 15 SP4, 15 SP5, and 15 SP6

Known issues

In this release, the OpenCL compiler enforces stricter type conversion checks. Kernel code that implicitly converts a global pointer to an integer without an explicit cast may now fail to compile with an “incompatible pointer to integer conversion” error. Most applications are unaffected, but if you encounter this error, update your kernel code to add an explicit cast or compile with the -Wno-error=int-conversion flag.

Features

General

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

Intel® C for Metal Compiler 

Disabled global fence on platforms other than Intel Data Center GPU Max Series.
Added 8-bit floating point conversion intrinsics.
Added a helper function to retrieve the global thread ID along its dimension.
Added support for new Battlemage and Panther Lake devices.
Updated the CM specification including the LSC memory interface, cache controls, and CM macro requirements.
Added the stochastic rounding intrinsic declaration.
Included the main <cm/cm.h> header implicitly, enabling caching when compiling from the CM source.
Added intrinsics for 2D block load and store operations.

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

Enabled the dynamic ICS via the opt-in KLV feature.
Updated the Graphics Micro Controller (GuC) to version 70.44.1.
Extended 2M userptr support to 1G.
Enabled backport support for kernel version 6.13.

Intel® Graphics Compiler for OpenCL™ 

Implemented GenISA predicated load/store intrinsics with promotion pass.
Added the ActiveThreadsOnlyBarrier option for OpenCL shaders.
Improved call site inlining heuristic.
Added a call merger pass that merges mutually exclusive function calls when they are too large to inline.
Added the __builtin_IB_disable_ieee_exception_trap and GenISA_disable_ieee_exception_trap intrinsic.
Introduced additional transpose block 2D SPIR-V APIs.
Added a flag to disable merging allocas of different types, providing better control over the merge alloca pass and disabling aggressive merging by default.
Added 32-bit ELF type support to ZeBin for x86 use cases.
Enhanced MergeAllocas performance by replacing all allocas, generating casts at the point of use, handling select instructions in liveness analysis, avoiding merging allocas across ContinuationHL calls in raytracing, and disabling allocas merging for raytracing.
Added support for recognizing OpenCL/SPIR-V built-ins represented as TargetExtTy to the ProcessFuncAttributes pass.
Enabled SIMD16 drop for Xe3 to minimize register spills.
Added the hasLscStoresWithNonDefaultL1CacheControls flag to zeinfo, enabling 3D clients to detect Load Store Cache (LSC) stores with non-default L1 cache policies for proper UAV coherency flushing.
Started using MergeAllocas for private memory merging, allowing reuse of non-overlapping private memory allocations to reduce overall memory usage.
Added support for SPIR-V MulExtended instructions to the Vector Compiler (VC).
Set the default General Register File (GRF) size to 128.
Added VISA support for HF8 conversion instruction and Panther Lake devices.
Enabled SetHasSample for the gather4* instructions.
Added support for stochastic round bf8 intrinsic in the Vector Compiler (VC).
Added Panther Lake support.
Added support for new Battlemage device IDs.
Introduced support for Floating-Point DIVide (FDIV) instructions inside IGCVectorizer.

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

Added support for the BUFFER_SIZE explicit argument.
Added the Level Zero API for querying kernel argument data.

Intel® Graphics Memory Management Library 

Improved handling of coherent and compressible resources.
Added support for new Battlemage device IDs.
Introduced the MOCS variable for Xe2.
Enabled GO:L3 for OpenCL usages.

Intel® ME TEE Library 

Integrated support for libmei version 1.6.4.
Added a C++ wrapper.
Introduced the TeeGetKind API.

Intel® Metrics Discovery Application Programming Interface 

Updated API version to 1.14.
Added support for offline metric calculation:
- OpenOfflineMetricsDeviceFromBuffer: Opens an offline metrics device object from a buffer.
- SaveMetricsDeviceToBuffer: Saves a metrics device to a buffer for future offline calculations.
- CloseOfflineMetricsDevice: Closes the offline metrics device object.
Added support in Xe KMD for configurable Overall Accuracy (OA) buffer size and half-full OA buffer interrupt handling.
Added new Battlemage device IDs.
Introduced the new EQUATION_ELEM_PREV_METRIC_SYMBOL equation element where prev$$"SymbolName" allows referencing the previous metric value within the local set.
Added support for the VectorEngine metric group.

Intel® Metrics Library for Metrics Discovery API 

Added support for Panther Lake.
Optimized copy query by reducing the number of GPU commands.
Added support for configurable Overall Accuracy (OA) buffer size in Xe KMD.
Improved performance of OA configuration updates.
Reduced build time.
Optimized debug helper.

Intel® oneAPI Level Zero 

Added support for registering a TeardownCallback to notify clients upon release of Level Zero resources.
Added support for sorting drivers based on provided devices.
Implemented basic leak checker in the validation layer.
Added zeImageViewCreateExt and zeMemFreeExt support to the leak checker.
Added API call logging to the validation layer.
Added the static Level Zero loader support.
Introduced support for 1.7 specification in the static loader.

Intel® Video Processing Library 

Introduced Intel® Video Processing Library API 2.15 support, including new property-based capability queries interface, extended decoder and encoder capabilities reporting, and definitions for VVC main 10 still picture profile and level 6.3.
Added the explicit INSTALL_EXAMPLES build option to control installation of example source code and content.
Updated the default Ubuntu build to version 24.04.

Intel® Video Processing Library GPU Runtime 

Improved AV1 decoding performance when all decode frame surfaces are in use.
Enabled property-based capability queries.

Intel® Video Processing Library Tools 

Introduced support for Intel® Media Transcode Accelerator.
Added new strings to the vpl-inspect tool to improve output readability.
Added the -props option to the vpl-inspect tool to support querying capabilities based on properties.
Updated the default Ubuntu build to version 24.04.

Improvements

Intel® Graphics Compiler for OpenCL™ 

Fixed a crash in GenerateBlockMemOpsPass that could occur in complex loops during analysis of memory access patterns when certain PHI node conditions were not met.
Fixed u8/u16 2D block read emulation by enforcing 4-column block width for alignment. Additionally, disabled decomposition of emulated d8/d16 transpose reads due to complexity and added emulation of d8/d16 transpose reads using native 2D block transform followed by mov instructions.
Fixed type arguments for creating the GenISA_sampleDCMlodptr call.
Removed unnecessary tracking and 32-bit truncation of bindless image offsets in kernel arguments, improving handling of bindless images and eliminating redundant instructions.
Started allocating the General Register File (GRF) number for Vector Compiler (VC) untyped load 2D intrinsics.
Fixed an issue with incorrect emission of phi values for structures.
Implemented a cycle-proof deletion strategy to Intel® Graphics Compiler Vectorizer now to ensure reliable cleanup when discarding vectorizer chains.
Fixed a crash in ProgramScopeConstantAnalysis that occurred during recompilation and resolved crashes encountered during the compilation of Blender kernels using the SPV_INTEL_bindless_images extension.
Set cache control for SPIR-V 2D block prefetch when cache control decoration is missing or invalid, provided the target device supports cache control.
Fix an iterator invalidation issue.
Fixed an issue with generating spill temporary variables for 4GRF operands.
Fixed segfault caused by default output stream flags in Intel® Graphics Compiler with SYCL by replacing them with printf.
Ensured consistent instruction order for PrivateMem to produce identical dumps in debug and ndebug compilations.
Started skipping debug calls in complex UnrollLoop loops.
Replaced std::map with llvm::MapVector.
Ensured uniform prefetch source address is GRF-aligned, enabled implicit arguments optimizations by default, and introduced GlobalOffset across Xe1, Xe2, and Xe3 platforms for improved performance and payload size reduction.
Added a check for the TotalGRFNum flag value before trying to return the value passed from an API option.
Started emitting bitcast after selecting a value in SimplifyConstant.
Stopped removing implicit kernel arguments, as they might be used by subroutines.
Started using the correct spill size for non-send destinations by rounding up to the nearest GRF size. Additionally, added a VISA option to enable spill cleanup within specified BB ID ranges to target transformations safely.
Started skipping dbg calls for vector aliasing heuristic.
Fixed potential nullptr dereference by passing the function as a reference to avoid null checks and improve efficiency.
Started reporting a warning when non-null/acc Architected Register File (ARF) register is used on ternary instruction.
Fixed issues related to opaque pointers support in GenXPacketizer.
Fixed issues causing incorrect chunk sizes in the ConstantCoalescing pass.
Optimized Read-Modify-Write (RMW) for strided first definitions in the entry basic block.
Moved VRT General Register File (GRF) bump-up after GRA optimizations.
Changed RayQueryDynamicRayManagement flag to be off by default due to stability issues. It can be enabled through Application Intelligence Layer (AIL).
Unified RayInfo between sync and async raytracing.
Fixed the fill checked built-in and implemented MAD built-ins for large shapes in the joint matrix.
Added a reserved VISA opcode and updated Intel Graphics Assembler (IGA).
Fixed a VISA assert issue.
Fixed a boundary condition issue.
Resolved an issue in the Vector Compiler (VC) affecting wrregion operations with bf16 source data types.
Fixed the lgamma_r behavior.
Fixed an issue where the atomic branch predicate was incorrectly selected when multiple modes were enabled, and removed the PreservesCFG flag from the InsertBranchOpt pass to avoid potential crashes.
Fixed initialization of PHI instructions of the i1 type.
Fixed handling the genx volatile pointer as a function argument in the Vector Compiler (VC).
Corrected the HWTID computation to use state registers when WMTP is unsupported by the shader type.
Stopped inserting the check/release intrinsics if the shader has discards.
Fixed the address register restriction in the destination.
Added support for fcvt with bf8 and hf8 data types.
Fixed the printf issues.
Fixed incorrect alignment in MergeUniformLoad with early return.
Fixed creation of chunk loads in the ConstantCoalescing pass.
Improved size reporting in payload_arguments in zebin.
Added a null check to prevent nullptr dereference in GenerateBlockMemOpsPass.
Fixed copying uniform variables.
Fixed the local copy propagation issue for indirect VxH source.
Stopped cloning debug instructions in CodeSinking and CodeLoopSinking passes.
Added conditional warning for dumped vector size, which is printed only when the ShaderDumpEnable flag is enabled.
Added correct predicate to the mov instruction when handling split samples.
Added lifetime.start emission for classic resource loops inside nested loops.
Enabled stateful rt stack for synchronized raytracing and separated sync and async rt stacks in raytracing magic types.
Initialized structure members to prevent potential nullptr dereference.
Fixed a DebugInfo issue in LLVM to avoid out-of-order evaluation.
Simplified the call to readFirstLanes for multiple getFirstLaneIDs.
Added support for functions with no return values and no arguments in SIMDCF.
Added a pass to remove freeze instructions prior to code generation.
Fixed handling rdregion operations with widths crossing register boundaries in the Vector Compiler (VC).
Corrected the maximum sub slices value in SIP.
Updated the execution mask to 32 bit on Xe2.
Corrected the sample_d_c and sample_d_c_mlod sampler message type.
Migrated ProgramScopeConstantResolution and StatelessToStateful to fix opaque pointers issues.
Disabled building legacy SPIR-V Reader.
Updated CopyVariableRaw to use SIMD32 as the maximum Single Instruction Multiple Data (SIMD) size on Xe2.
Deprecated the initial set of generation 9 Vector Compiler (VC) LITs, replacing them with XeHPG equivalents.
Made code assumption in get_global_id optional.
Extended the application of the multiplication pattern in GetElementPtr Loop Strength Reduction (LSR) pass to improve performance.
Added options for controlling the depressurizer thresholds.
Removed too strict restrictions from the LICM pass.
Restored the SIMD16 drop functionality on Xe2, enabling support for spilling kernels using SIMD16 on this architecture.

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

Updated the COMPUTE_WALKER to fix incorrect RawData array length.
Added the FillImage1dBuffer built-in kernel.
Started blocking zeContextMakeImageResident.
Started failing device initialization if kernel debugging is misconfigured, with a detailed error message printed to stderr.
Started passing the Deallocate2 callback to the Graphics Memory Manager (GMM).
Corrected the Xe sysfs paths for the Compute Command Streamer (CCS) mode setting.
Made external semaphore controller thread-safe and ensured proxy events are destroyed only when the semaphore thread controller releases resources to prevent sporadic failures.
Improved ULLS light ring handling by managing new ring buffer residency, extending mutex protection for safe stop operations, and updating USM cleaner to properly stop ULLS light during resource eviction.
Ensured Zebin is dumped during program build when unpackSingleDeviceBinary is not called, provided the debug key is enabled.
Corrected gfx_core_helper definitions for EUSS.
Correct logic for retrieving valid timestamp bits.
Removed overflow check in calculations for Xe2+ cores using EUSS.
Improved media engine handling
Started returning an error via paraminfo if it is queried with a parameter count of 0, but the programmable actually has one or more parameters.
Started patching payload arguments in inline data in case of indirect kernels.
Ensured payload arguments are patched before fetching the walker command.
Added initial support for single temporary allocations list and ensured flush of split task count.
Replaced sfence with mfence on discrete devices and moved ULLS semaphore to shared memory on Xe2.
Disabled deferring Memory Object Control State (MOCS) on WSL for Lunar Lake and unified deferring MOCS to the Page Allocation Table (PAT).
Updated implementation to expose the THREAD_SCRATCH debug register only when running in the heapless mode.
Improved cache handling by invalidating texture and heap caches before reuse or image reads.
Improved compression handling by disabling it for pre-Xe2 platforms, enforcing capability table flags, and removing a redundant workaround for Alchemist GPUs.
Added input/output control helper for context destruction.
Changed the stype member type in Level Zero Core and Tools driver extensions to uint32_t-alias to prevent casting outside the ze_structure_type_t/zet_structure_type_t enum range
Improved fence allocation and synchronization by passing product helper to isFenceAllocationRequired, using global fence helper, simplifying fence selection in ULLS, and removing global fence from CW post-synchronization on Battlemage.
Improve container management by reserving residency before addition, unifying non-append method calls, and preventing queue buffer consumption during command list execution.
Enabled staging infrastructure for 3D images.
Updated semaphoreBuffer and ringBuffer usage on integrated devices.
Improved the NonCopyableOrMovable and NonCopyable concepts.
Unified the local memory size getter for i915 and Xe.
Set the vmbind user fence in makeMemoryResident to resolve the memory reporting issue.
Corrected the allocation size in freeSVMAlloc to prevent crashes.
Added an option to enable and disable the heapless mode in the OpenCL offline compiler.
Parsed the Compute Command Streamer (CCS) mode setting for platforms other than Intel Data Center GPU Max Series.
Implemented per-element BLT copying for tiled 1D arrays and began treating tiled 1D images as 2D with a height of 1 for BLT operations.
Stopped enabling compression on xe_lpg for Linux and WSL.
Corrected blit properties for CL_MEM_OBJECT_IMAGE1D_BUFFER images.
Corrected the ZE_MEMORY_ACCESS_CAP_FLAG_CONCURRENT reporting.
Enabled the Unified Shared Memory (USM) compression on Linux.
Aligned allocation sizes of 2MB or larger in local memory to a 2MB boundary. This behavior is controlled by the is2MBLocalMemAlignmentEnabled function. Additionally, implemented a pool allocator for gpuTimestampDeviceBuffer. Allocations are shared per device and controlled via the EnableTimestampPoolAllocator debug flag.
Corrected allocation in MemObj::getMemObjectInfo and created graphicsAllocation per each rootDevice.
Corrected Level Zero versioning.
Corrected a Device IDs mismatch.
Started passing the ReadOnly flag only for page-misaligned input pointers.
Corrected the returned metric value counter for EU stall.
Fixed an issue causing memory allocation crash.
Fixed a shared memory failure issue.
Fixed a hang issue in the Toolbox Interface (TBX) page fault manager.
Implemented changes to prevent race conditions during resource eviction.
Configured scratch pages for the debugger.
Corrected allocation handling for the increment Command Buffer (CB) event.
Enabled support for image array types with an array size of 1 on Xe2 and later platforms.
Corrected the order of passing arguments to obtainCommandStream.
Added a check for the Shared Virtual Memory (SVM) allocated host pointer in clCreateBuffer.
Merged preliminary and non-preliminary code for the Legacy Sysman Memory module.
Removed the patchtoken fallback.
Corrected the error code for the deprecated clSetCommadQueueProperties.
Removed operation access for unsupported types.
Fixed an issue with setting up the Compute Command Streamer (CCS) mode.
Corrected literal raw strings handling in the printf formatter.
Added support for MetricCreateFromProgrammableExp2.
Corrected DSH generation and programming of inline samplers with bindless addressing in Level Zero.
Started passing the root device when creating secondary contexts to ensure proper initialization of gfxCoreHelper in Direct Rendering Manager (DRM).
Enabled 2-way coherency for misaligned user memory.
Preserved the allocation type for memory objects.
Added support for passing -device and -device_options in multiple formats in the OpenCL Offline Compiler.
Fixed the status return value in getExternalMemoryProperties when operating in the Toolbox Interface (TBX) mode.
Removed the deprecated LayoutRight Graphics Memory Manager (GMM) flag.
Introduced the ImageSurfaceState helper class and relocated global functions into the class to reduce compilation time.
Added the Sysman device directory name as a parameter to SysmanKmdInterface.
Set the external semaphore version in Level Zero.

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

Fixed an issue where the sched_setattr_nocheck API was not exported in kernel versions earlier than 5.14.
Switched to locked variant of wake_up_interruptible for safer thread wake-ups.
Reduced spurious wake-ups for single-task shmem/userptr jobs.
Started propagating wake-up from suspended threads to avoid delayed task execution.
Replaced function type casting with typed function stubs.
Added a reference around vm_bind to maintain the Virtual Memory Area’s (VMA) validity.
Started clearing the Multi Die Fabric Interconnect (MDFI) boot time errors, as they are expected during the initialization of MDFI fabric and may be confused with runtime errors.
Started using the kobject attribute instead of the device attribute for num_cslices and ccs_mode sysfs entries on RHEL 8.X.
Started handling additional PCI AER corner cases to be able to reset devices without locking up the machine.
Reordered hardware waits and GPU reset logic during PCI faults to avoid blocking on unresponsive hardware while recovering from a hardware failure.
Fixed an issue where a mutex could be held indefinitely when attempting to remove an idle Virtual Memory Area (VMA) from the VM.
Prevented memory allocations during page faults triggered by GPU reset.
Updated GTT_MMAP_VERSION to align with corresponding changes in user space.
Allowed data to be discarded on forced unbinds, avoiding swaps to inaccessible system memory.
Set the lmem_offset to 0 after use so that the next local memory block does not carry the same offset leading to lost data during Single Root I/O Virtualization (SR-IOV) migrations.
Fixed incorrect annotations.
Fixed error unwinding in i915_virtualization_probe.
Added periodic checks for forward progress by monitoring context switches and user interrupts. If the same context remains active without interrupts since the last check, a warning is generated with no further action.
Prevented default context creation when wedged.
Cleaned up faulting initialization.
Prevented DPC NPD after initialization failure by early iaf setup and driver-device decoupling on probe failure.
Started protecting per-CPU px_cache from interrupts.
Started sending a TLB invalidation request after each Virtual Memory Area (VMA) binding for GuC use, instead of deferring until before enabling GuC, to prevent Single Root I/O Virtualization (SR-IOV) failures.
Started periodic check for mmio failures.
Started handling CT fault injection during early initialization by ensuring CT descriptor objects are not dereferenced before assignment, preventing failures on early faults.
Started checking for context creation failure during execbuf.
Added support for deferred context attachment to existing clients.
Removed the residual calls to the empty i915_oa_init_reg_state to completely excise an old use-after-free.
Skipped the HuC authentication register check as it is no longer needed.
Prevented soft lockup during defragmentation on eviction.
Prevented a potential compute hang on Alchemist GPUs.
Updated CT desc->head after consuming a receive chunk to prevent buffer overflow and slow GuC messaging.
Added device PCI IDs to GPU dumps.
Updated ce->vm on parallel child contexts.
Corrected the CSC hardware errors.
Added the eudbg event for deferred default context allocation.
Removed lockdep assertions around Global Graphics Translation Table (GGTT) updates to prevent conflicts.
Preserved Translation Lookaside Buffer (TLB) seqno when splitting clear pages into multiple smaller pages if there is an outstanding TLB invalidation for those pages.
Deferred the default context allocation until first use, reducing overhead when a device opens.

Intel® Graphics Memory Management Library 

Disabled compression for the GMM_FORMAT_I420 format.
Added size validation when checking NoOptimizationPadding.
Enforced the Tile4 layout over Linear for flipchain resources.
Resolved type incompatibility issues.

Intel® ME TEE Library 

Reduced hardware register polling timeout in EFI.
Refactored HECI_DEVICE_KIND handling in EFI for better maintainabilit.
Fixed an EFI issue by changing the propertyMap array type to CHAR8*.
Fixed EFI compilation errors with GCC.

Intel® Metrics Discovery Application Programming Interface 

Fixed incorrect bitfield parsing in metric equations.
Corrected scaling of std.color.node_pbe_arb metric on Lunar Lake.
Optimized memory allocation size to improve performance and reduce overhead.
Updated the OpenIoStream behavior to return CC_ERROR_NOT_SUPPORTED when processId != 0.
Removed legacy global symbols to reduce namespace clutter and improve maintainability.

Intel® oneAPI Level Zero 

Fixed Sysman-only initialization to prevent retrieval of the loader context when version compatibility is not met.
Corrected version and GUID updates for version 1.22.2.
Fixed GUID generation and updated to version 1.22.3.
Resolved an issue in zesInit to correctly initialize the requested API version.
Fixed artifact upload workflow.
Fixed the extension validation logic.
Improved initialization error checking to verify validation layer behavior.
Fixed experimental extension validation to accept unknown extensions within a valid range.
Corrected sType initialization in property query operations.
Improved teardown checks to prevent invalid context usage.
Added the missing header to ze_ddi_common.h.
Fixed enabling the Digital Display Interface (DDI) handle extensions.
Fixed the incorrect sType assignment in zello_world.
Modified context_t to always allocate dynamically and support delayed destruction.

Intel® Video Processing Library 

Updated the model used in the interop example to a vehicle detection model.
Fixed the BUILD_EXAMPLES build option so it no longer depends on INSTALL_DEV to take effect.
Removed outdated Docker files provided with examples.

Intel® Video Processing Library GPU Runtime 

Resolved crash issues during AV1, AVC, and HEVC decoding related to surface creation on resolution changes.

Intel® XPU Manager and XPU System Management Interface 

Improved the firmware update under the recovery mode for Intel® Data Center GPU Flex Series.
Introduced security enhancements.

2025-04-29

The 2507.17 release supports the following operating systems:

Red Hat Enterprise Linux (RHEL): 8.8, 8.10, 9.2, 9.4, and 9.5
Ubuntu 22.04 and 24.04
SUSE Linux Enterprise (SLES): 15 SP4, 15 SP5, and 15 SP6

Improvements

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

Updated the Graphics Micro Controller (GuC) to version 70.44.1.
Resolved a hang detection issue on Intel Data Center GPU Max Series by re-enabling GPU hang checks. Hang detection now only logs a warning message without terminating the application.

Intel GPU Firmware 

Updated the Graphics Micro Controller (GuC) to version 70.44.1.

2025-03-18

The 2507.12 release supports the following operating systems:

Red Hat Enterprise Linux (RHEL): 8.8, 8.10, 9.2, 9.4, and 9.5
Ubuntu 22.04 and 24.04
SUSE Linux Enterprise (SLES): 15 SP4, 15 SP5, and 15 SP6

Features

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) 

Added support for the HBM_REPLACE bit to signal High Bandwidth Memory (HBM) health status and its transition to the REPLACE state. This enhancement enables the driver to detect the bit and prevent loading when the state changes to REPLACE, while also reporting the issue and prompting HBM replacement.

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

Started handling page fault events in the Xe debugger.
Added support for the cl_khr_expect_assume OpenCL extension that introduces mechanisms to supply the compiler with information that can enhance the performance of certain kernels.
Implemented the Level Zero zeKernelGetBinaryExp API that allows retrieving kernel binary program data.
Added support for shared system Unified Shared Memory (USM) allocation in appendLaunchKernel.
Implemented enhancements to the Unified Shared Memory (USM) reuse mechanism, including the introduction of a USM reuse cleaner that efficiently manages system and local memory across different reuse strategies, as well as an extension of the USM reuse limit infrastructure.
Improved cache management by supporting whitelisted includes.
Added support for handling new Reliability, Availability, and Serviceability (RAS) errors in Sysman.
Implemented alignment of host Unified Shared Memory (USM) to 2MB on discrete devices when the allocated size exceeds 2MB.

Intel® Graphics Compiler 

Modified the pass threshold to optimize the i64 multiplication performance.
Introduced Panther Lake support.
Improved vectorizer to support vector emission of ftrunc instructions.
Enabled the IndVarSimplification pass to improve performance.
Enabled access to the Workload Management and Thread Programming (WMTP) SIP kernel for the Xe3 core and introduced a default WMTP SIP configuration per Shared Local Memory (SLM) for Xe2.
Added more aggressive late rescheduling phase to the CodeLoopSinking pass and an option to disable the maximum sinking heuristic in the presence of 2D block reads.
Improved the InlineHelper LLVM utility.
Implemented the MergeAllocas pass and enabled allocation merging prior to the split asynchronous pass.
Enabled the emission of vectorized floating-point addition (FADD) instructions, allowing the VISA emitter to process them efficiently.
Implemented nested 3D resource loop unrolling.

Intel® oneAPI Level Zero 

Upgraded specification to version 1.12.15.

Intel® Video Processing Library 

Introduced support for Intel® VPL API 2.14, introducing new quality and speed settings for AI-powered video frame interpolation. This update also expands algorithm and mode selection options for AI-based super resolution and adds support for High Efficiency Video Coding (HEVC) level 8.5 decoding.
Improved compatibility with Python 3.12 development environments.

Intel® Video Processing Library Tools 

Integrated screen content coding tools for AV1 into sample_encode.
Added a new GTK renderer option to sample_decode and sample_multi_transcode.
Introduced a new -fullscreen option for GTK in sample_decode and sample_multi_transcode. Users can now toggle full screen using Ctrl+f and exit with Esc.
Enhanced support for Python 3.12 development environments.

Changes

General

Updated the signing key for KMD prebuilds to enhance security and ensure continued reliability. The new key, valid for one year, will be used to sign all new releases. To ensure compatibility with these updates while maintaining the secure boot functionality, you need to download and install a new Distinguished Encoding Rules (DER) certificate.

Intel® Graphics Compiler 

Lowered bfloat ceil and floor intrinsics.
Refactored parameters in vc-lits in lit-config for LLVM 16 to not link the initializeGenX function.
Increased the early recompilation threshold for default General Register File (GRF) to 500.
Enabled the EnableWaveShuffleIndexSinking registry key by default.
Enabled the WaveAllJointReduction pass by default.
Added an extra assertion check to the SIMDInfo offset.

Improvements

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

Introduced page fault handling improvements.
Fixed an issue causing the CSC hardware errors.
Removed unnecessary lockdep debugging checks from Global Graphics Translation Table (GGTT) updates.
Fixed timeout issues by preserving Translation Lookaside Buffer (TLB) seqno when splitting clear pages.
Fixed issues causing compilation errors on kernel 6.6 and later.
Fixed an issue where prefetch was attempted on empty objects.
Fixed an issue where pid_task() could fail if the target process had already exited.
Implemented a workaround for Address Translation Services for Memory (ATS-M) and introduced support for G8 power state to reduce idle power consumption.
Modified the logic to avoid calling pm_qos_request a second time on an existing request during breadcrumb reset.
Disabled C-states for breadcrumb interrupts to reduce Direct Memory Access (DMA) latency.
Cleaned up incomplete shmemfs obj->base.filp on failed swapout.
Hardcoded memory health status in sysfs to prevent breakage.
Implemented flushing of freed objects before reporting available memory to stabilize the reported memory levels.
Modified implementation to retry eviction only when it is blocked by active or locked objects, aiming to reduce response time.
Optimized Virtual Memory Area (VMA) prefetch by short-circuiting redundant operations.
Corrected Compressed Color Surface (CCS) copies for Single Root I/O Virtualization (SR-IOV) save and restore.
Restricted shmem flags to a valid set for swapin to resolve a page fault issue.
Modified the implementation to repeat the Translation Lookaside Buffer (TLB) flush invalidation request, resolving the issue with the failing Hardware Performance Library (HPL).
Removed early unlocked unbind from object free to avoid race conditions between lockless unbinding and eviction of non-persistent VMAs.
Introduced changes to protect i915_drm_client_fini from early shutdown.
Started supporting compilation with CONFIG_PAGE_TABLE_ISOLATION to fix a compilation issue on RHEL.
Optimized the unbind step in the GT IFR flow by skipping context runtime updates when the device is quiesced. This change reduces the execution time.

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

Disabled implicit callback conversion for wait events to resolve the wait operation hang issues.
Added the missing callback event cache flush to fix an issue with the zeEventHostSynchronize hangs.
Fixed an issue with reporting EU counts for multi-slice platforms.
Fixed an issue where ZE_AFFINITY_MASK was not working when ZE_FLAT_DEVICE_HIERARCHY was set to COMBINED in OpenCL.
Implemented shared allocations to preserve reference timestamps and introduced a flag in the Inter-Process Communication (IPC) pool data to verify if the mapped timestamp flag is set.
Fixed an issue where event_profiling::command_start returned an incorrect result.
Set stateless addressing mode for buffers that are neither bindful nor bindless.
Started retrieving the minimal offset size for region barrier.
Fixed the scope of the result variable in initDriver to resolve an issue where it was defined in a narrower scope, causing the initialization result to be improperly discarded.
Started returning rawDataSize as zero when the readIoStream call fails.
Resolved issues with parsing and setting the Level Zero debugger bitmask.
Fixed performance issues on Battlemage GPUs.
Ensured memory residency by setting the vmbind user fence when making memory resident.
Prevented crashes due to over-allocation by introducing a defer backing flag to Graphics Execution Manager (GEM) create input/output control, ensuring memory is resident before locking.

Intel® Graphics Compiler 

Fixed regression chart dump for General Register File (GRF) configurations with more than 128 registers.
Resolved issues in the vectorizer, improving its stability and performance.
Disabled the TrivialLocalMemoryOpsElimination pass from the pipeline.
Resized G4_Declare’s row size for atomic operations to prevent out-of-bounds (OOB) issues.
Fixed issues in VISA parser for Load Store Cache (LSC) 2D block operations to allow mixed register and immediate AddrX and AddrY operands for 2D block load and store instructions.
Fixed value tracker handling of Global Element Pointers (GEP) with zero indices by treating them as bitcasts to prevent confusion in kernel usage.
Implemented dynamic optimization threshold adjustment for the depressurizer based on the number and size of General Register File (GRF) registers.
Fixed incorrect instructions placement in the rollback functionality of the CodeLoopSinking pass.
Stopped using the %sp and %fp predefined variables.
Implemented dedicated logic for handling discards in DynamicRayManagementPass to prevent crashes.
Fixed the direct address destination restriction on SIMD32.
Improved the alignment calculation in constant coalescing and started supporting additional load and store intrinsics in SynchronizationObjectCoalescing.
Fixed incorrect condition check in isRegionInvariant of WIAnalysis.
Stopped removing Built-in Function (BiF) module prebuilt stamp files to avoid redundant recompilations when CMake files are updated.
Fixed an issue with constant folding prevention inside loops.
Added intrinsic cache to KernelDebugInfo and prevented indirect access to Software Scoreboard (SWSB).
Reduced the number of atomics hitting the same cache line by performing atomic predication.
Enabled the Execution Out-of-Order Thread (EOT) to participate in the Software Scoreboard (SWSB) token assignment.
Fixed a channel mask issue in the src0 length for RenderTargetDataPayload, where the Alpha channel was incorrectly controlled.

Intel® oneAPI Level Zero 

Addressed and fixed potential memory leaks.
Fixed issues in the generation of pkg-config files.
Corrected code generation for libddi table queries.
Corrected validation layer’s parameter checker for extensions.

Intel® Video Processing Library 

Fixed the bootstrap process to support Debian distributions that do not define the ID_LIKE property.

Intel® Video Processing Library Tools 

Fixed the bootstrap process to support Debian distributions that do not define the ID_LIKE property.