Intel® oneAPI Math Kernel Library (oneMKL) Release Notes and System Requirements

Where to Find the Release

You can get the latest version of Intel® oneAPI Math Kernel Library (oneMKL) as part of the Intel® oneAPI Toolkit, or as a Stand-Alone download.

If you have Priority Support for Intel® oneAPI Toolkit or any other oneAPI toolkit (i.e. an unexpired free license or a paid license with active support at the time of the build date for this product release), you can download the installer by logging in to Intel® Registration Center and selecting the appropriate product. You may need to create an account and/or register your product. For additional information please see: Product Registration and Sign-Up FAQ

New in this Release

2026.2

Highlights

• SYCL link line in CMake now uses OpenMP* as the default threading layer. 
• SYCL buffer APIs will now throw deprecation warnings. 
• Some Sparse BLAS SYCL routines now support half (fp16) precision. 
• Inspector-Executor Sparse BLAS routines now support verbose mode. 
• Floating point exceptions have been significantly reduced. 
• Improved performance for Random Number Generators on Intel® CPUs with Intel® AVX10 ISA available.

Improvements and Changes

• BLAS 

  • Features

    • SYCL graph recording and replay is supported for select BLAS functions.
    • A “gemmtr” routine is introduced to replace “gemmt” and align with the BLAS standard.

  • Optimizations

    • Improved the performance of nrm2 and iamax for double and complex double precisions.

• Sparse BLAS

  • Features

    • Select support added for sycl::half (fp16) data type in Sparse BLAS SYCL USM APIs targeting CSR format (sparse::set_csr_data/release_matrix_handle) and sparse::gemv/trmv/gemm/omatcopy/sort_matrix operations.   
    • Inspector-Executor Sparse BLAS has added support for modern Fortran APIs with 64-bit integers (matching previously added C APIs in oneMKL 2024.2) for both lp64/ilp64 interfaces using the *_64 tag in the API name and MKL_INT64 as the integer type. 
    • Inspector-Executor Sparse BLAS TRSV/TRSM APIs now support unsorted input matrices in all formats (previously only guaranteed success for sorted input matrices). 
    • Verbose mode is now supported for all C/Fortran Inspector-Executor Sparse BLAS routines on CPU.
      • As in other domains, it can be enabled through the MKL_VERBOSE environment variable or by calling mkl_verbose(int mode), and it reports: 
        • a trace of the computational functions that are called, 
        • the parameters passed to each function and their values, and 
        • the execution time of each function. 
      • Unlike other domains, Inspector-Executor Sparse BLAS emits two lines per function call. The first line begins with MKL_VERBOSE BEGIN and reports the function name and its parameters. The second line begins with MKL_VERBOSE END and reports additional details such as execution time and, in threaded builds, the thread ID and number of threads. Future releases may add further information between the BEGIN and END lines to describe the internal steps of the computation. 

  • Optimizations

    • Improved performance for Inspector-Executor Sparse BLAS sparse matrix multiplication routine mkl_sparse_sp2m() on CPUs for CSR/CSC/BSR formats, especially when using multistage symbolic/numeric breakdowns, in the numeric stage (using request SPARSE_STAGE_FINALIZE_MULT)  when solving for the matrix values of output matrix, when the output sparsity pattern has previously been computed (using request SPARSE_STAGE_FULL_MULT_NO_VAL, or SPARSE_STAGE_ NNZ_COUNT + SPARSE_STAGE_FINALIZE_MULT_NO_VAL). 

• LAPACK

  • Optimizations

    • Improved the single-thread performance for LU factorization, getrf.

• DFT

  • Features

    • Introduced new `DFTI_CONFIG_PARAM` enum values, namely `DFTI_FWD_DISTANCE` and `DFTI_BWD_DISTANCE`, for setting the transform distances between batches in the forward and backward domains respectively. Similarly, `DFTI_FWD_STRIDES` and `DFTI_BWD_STRIDES  were introduced to set the strides along the transform dimensions in the forward and backward domains, respectively.

• Vector Math

  • Optimizations

    • Implemented Intel® AVX10.2 optimizations for the oneMKL Vector Math Library, leading to performance improvements for vector math functions on platforms with Intel® AVX10.2 support.

• Vector Statistics 

  • Optimizations

    • Improved performance for MCG59, MRG32K3A, MT2203, ARS-5, PHILOX4X32X10 engines on Intel® CPUs with Intel® AVX512 ISA available. 

• Sparse Solvers 

  • Optimizations

    • New options for matching and scaling can improve performance of the reordering phase of Pardiso for some matrices.
    • Performance of dcsrilu0 preconditioner has been greatly improved and threading support has been added.

Fixed Issues

• Fixed potential integer overflow in mkl_get_env_int and mkl_mpi_get_ppn APIs. 

• Fixed clobbering warnings by avoiding file conflict at install time for conda packages. 

• Fixed a performance degradation for some BLAS GEMM operations on Intel® Data Center GPU Max Series.

• Fixed several bugs that caused the Pardiso Schur Complement feature stability issues.

• oneMKL now can throw proper exception in case of unknown GPU architecture.

Support Deprecated

• BLAS gemmt routine is deprecated, use gemmtr instead. 

• DFT `DFTI_CONFIG_PARAM` values `DFTI_INPUT_STRIDES`, `DFTI_OUTPUT_STRIDES`, `DFTI_INPUT_DISTANCE`, and `DFTI_OUTPUT_DISTANCE` are deprecated in favor of the new values `DFTI_FWD_STRIDES/DFTI_BWD_STRIDES` and `DFTI_FWD_DISTANCE/DFTI_BWD_DISTANCE`; support for the deprecated values will be removed in a future release. 

2026.1

Highlights

• Introduced initial support for Intel® GPUs that are based on Intel® Xe3P microarchitecture.

Improvements and Changes

• LAPACK

  • Optimizations

    • Improved the performance of lapack::orgqr for all architectures of Intel® GPUs.

Fixed Issues

• Fixed potential out-of-bounds read in mkl_verbose_output_file Fortran API. 

Support Deprecated

• mkl_set_memory_limit API and MKL_FAST_MEMORY_LIMIT environment variable are deprecated. Support will be removed in future releases. 
• oneMKL SYCL Buffer APIs are deprecated for all domains. Support will be removed in future releases. 

2026.0

Highlights

• Introduced performance optimization for MCG59 engine on Intel® Xeon® 6 processor with P-cores. 
• Intel® Compiler’s qmkl option behavior for oneMKL SYCL* APIs is aligned with the C/Fortran APIs to link with the LP64 interface (Intel® Compiler release notes). 
• Hardened all oneMKL binaries with Intel® CET and all C libraries with FORTIFY_SOURCE. 
• oneMKL SYCL and OpenMP* offload API support system shared USM pointer on Linux*. 

Improvements and Changes

• BLAS

  • Optimizations

    • Improved the performance of nrm2 on Intel GPUs. 
    • Performance and threading improvements for many level-2 BLAS functions.
    • Improved performance of iamax and iamin level-1 BLAS functions. 

• LAPACK 

  • Optimizations

    • Improved the performance of generalized eigensolver routines {s,d}ggev for both single-threaded and multi-threaded use cases. 
    • Improved the performance of QR factorization routines, geqrf and geqrf_batch (strided), for Intel® Arc™ A-Series Graphics and Intel® Data Center GPU Flex Series. 
    • Improved the performance of getrf_batch (group) for small and large group counts on Intel® Data Center GPU Max Series. 

• DFT

  • Features 

    • Enabled support for non-default slab and custom distribution for the SYCL distributed DFT APIs on Intel® GPUs.

  • Optimizations

    • Improved the performance of power-of-two 1D complex FFT with medium size on Intel® Core™ Ultra Processors (Series 3). 
    • Improved the performance of complex and real large-sized FFT on Intel® Arc™ B-series Graphics, Graphics for Intel® Core™ Ultra Processors (Series 2 and Series 3).

• Vector Math

  • Features 

    • Intel® Advanced Vector Extensions 10.2 (Intel® AVX10.2) implementations were added with corresponding architecture dispatch sequence. 

• Vector Statistics

  • Optimizations

    • Improved performance for MCG59 engine on Intel® CPUs with Intel® AVX512 ISA available. 

• Sparse Solvers 

  • Features 

    • Extended Eigensolvers: Introduced an option for computing the total number of eigenvalues within an interval. If fpm[49] for C/C++ programs or fpm(50) for Fortran programs is set to 1, Extended Eigensolver routines for CSR format compute the total number of eigenvalues. 

  • Optimizations

    • oneMKL Pardiso has significantly improved scalability on multi-core systems.
    • The performance of oneMKL Pardiso is improved in Conditional Numerical Reproducibility (CNR) mode. 

• Library Engineering

  • Optimizations

    • Introduced new AVX10.2/APX optimizations as preview code-path. Use the environment variable MKL_ENABLE_INSTRUCTIONS=AVX10 or the related API to enable it.

Fixed Issues

• Fixed the issue that File Patch Version always shows “1” on Windows*. 

• Re-enabled some Lapack _64_ Fortran symbols that were missing in the interface libraries since oneMKL 2025.2. 

• Various stability improvements in oneMKL Pardiso and ScaLAPACK. 

• Fixed floating-point exception raised by DftiCommitDescriptor on Intel® AVX512 ISA. 

Support Deprecated

• Intel® SSE4.2 specific optimizations are deprecated and will be removed in future releases. 
• oneMKL Link Tool’s interactive mode is deprecated and will be removed in future releases. 
• The Convolution and Correlation AUTO mode computation is deprecated and will be removed in future releases. Use the FFT mode instead or evaluate performance for your use case of the DIRECT and FFT modes (DIRECT mode might be faster for really small sizes).

Support Removed

• The oneMKL static SYCL library is deprecated from the 2025.2 release and will be removed in the 2026.0 release. Users are encouraged to use the oneMKL dynamic SYCL domain-specific libraries instead. For guidance on linking against oneMKL, please refer to the Link Line Advisor. 
• oneMKL vars scripts will stop appending CPATH starting from the oneMKL 2026.0 release. IFORT users will need to append CPATH manually if required. Other users will not be affected. 
• oneMKL will not be released on the Cloudera distribution channel starting oneMKL 2026.0. 
• Most NIST-like Sparse BLAS Level 2 and Level 3 APIs have been removed from oneMKL product as of the 2026.0 release. The remaining supported NIST Sparse BLAS APIs are mkl_?csradd, mkl_?csrmultcsr, and mkl_?dnscsr. Users are advised to transition to the Inspector Executor Sparse BLAS routines for sparse matrix operations. For conversions between sparse and dense formats, it is recommended to use the newer conversion APIs introduced in oneMKL 2025.3.  
• The SYCL DFT APIs were marked as deprecated in the oneMKL 2025.0 release or earlier have been removed. 
• Random number generation save_state/load_state API with std::string as a second parameter have been removed. 

Known Issues and Limitations

2026.2

• Intel® MPI 2021.17 is not supported on Ubuntu 26.04. Use newer Intel® MPI versions for this Ubuntu version or use older Ubuntu with Intel® MPI 2021.17. Full system requirements for Intel® MPI can be found at Intel® MPI Library Release Notes for Linux* OS. 

• oneMKL BLAS imatcopy_usm and omatcopy_usm routines may fail on Graphics for Intel® Core™ Ultra Processors (Series 2) on Windows* OS.

• FEAST sparse eigensolver routines may crash on Windows* when built with /guard:cf option. The recommended workaround is to omit this option.

• oneMKL BLAS functions for bfloat16 precision may fail on Graphics for Intel® Core™ Ultra Processors (Series 2).
• Performance regressions for some very small sizes of oneMKL CGEMM (only single-precision complex) may be observed on CPU.
• oneMKL DFT may give incorrect results for some large real problems on CPU. 
• Selected oneMKL SYCL functionality requires OpenCL* runtime when using the Level Zero SYCL backend on Intel® GPUs. 
• oneMKL SYCL DLL could leak memory after unloading on Windows*. The problem can be avoided by adding mkl_free_buffer() before unloading the DLL. 
• oneMKL FFT with a large prime factor (greater than 1024) may fail on Intel® Data Center GPU Max Series. 
• oneMKL DFT SYCL APIs don’t support negative strides and distances.

2026.1

• Vector Math Device SYCL API may fail to execute on Windows because of missed mkl_sycl_vm.6.dll. The problem can be avoided by explicit loading this DLL as shown below: 
  // at the top of the file 
  #include <windows.h> 
  // at the beginning of the main function 
  HMODULE H = LoadLibraryA("mkl_sycl_vm.6.dll"); 

• oneMKL PARDISO solve phase may hang in some cases for large matrices when using TBB threading. OpenMP* threading is recommended. 

• oneMKL DFT SYCL APIs using SYCL buffer for data input do not support SYCL sub-buffer inputs for large power-of-two sizes [2²¹,2²⁶] 1D complex FFT. 

• oneMKL FFT with a large prime factor (greater than 1024) may fail on Intel® Data Center GPU Max Series. 

• oneMKL DFT SYCL APIs don’t support negative strides and distances.

• oneMKL LAPACK direct call routines on Windows* with TBB threading may crash. OpenMP* threading is recommended.

• oneMKL PARDISO pivoting phase may fail sporadically on hardware not running Intel® AVX2 and later code paths in rare cases.  

• oneMKL Sparse BLAS use of sycl::buffer version of sparse::update_diagonal_values() on the Linux* operating system may sporadically cause unexpected crashes on Intel® Iris® Xe Max Graphics, Intel® Arc™ Graphics and Graphics on Intel® Core™ Ultra Processors (Series 2). 

• oneMKL Sparse BLAS use of the sycl::buffer version of sparse::matmat() API on Windows* may sporadically produce incorrect results on Graphics on Intel® Core™ Ultra Processors (Series 3). 

• Selected oneMKL SYCL functionality requires OpenCL* runtime when using the Level Zero SYCL backend on Intel® GPUs.

• oneMKL SYCL DLL could leak memory after unloading on Windows*. The problem can be avoided by adding mkl_free_buffer() before unloading the DLL. 

2026.0

• oneMKL DFT SYCL APIs using SYCL buffer for data input do not support SYCL sub-buffer inputs for large power-of-two sizes [2²¹,2²⁶] 1D complex FFT. 

• oneMKL FFT with a large prime factor (greater than 1024) may fail on Intel® Data Center GPU Max Series. 

• oneMKL DFT SYCL APIs don’t support negative strides and distances.  

• oneMKL LAPACK direct call routines on Windows* with TBB threading may crash. OpenMP* threading is recommended. 

• oneMKL PARDISO pivoting phase may fail sporadically on hardware not running Intel® AVX2 and later code paths in rare cases. 

• oneMKL FEAST extended eigensolver may fail on Windows with TBB threading on hardware not running Intel® AVX2 and later code paths. OpenMP* threading is recommended. 

• oneMKL Sparse BLAS use of sycl::buffer version of sparse::update_diagonal_values() on the Linux* operating system may sporadically cause unexpected crashes on Intel® Iris® Xe Max Graphics, Intel® Arc™ Graphics and Graphics on Intel® Core™ Ultra Processors (Series 2). 

• oneMKL Sparse BLAS use of the sycl::buffer version of sparse::matmat() API on Windows* may sporadically produce incorrect results on Graphics on Intel® Core™ Ultra Processors (Series 3). 

• Selected oneMKL SYCL functionality requires OpenCL* runtime when using the Level Zero SYCL backend on Intel® GPUs. 

• oneMKL SYCL DLL could leak memory after unloading on Windows*. The problem can be avoided by adding mkl_free_buffer() before unloading the DLL. 

• Data fitting cubic splines interpolation SYCL API may produce incorrect results on Intel® GPUs. 

• 3rd Central Sum and Central Moment output of Summary Statistics SYCL and OpenMP* offload routines may produce incorrect results on Intel® Arc™-A Series and Graphics on Intel® Core™ Ultra Processors (Series 2). 

• Certain BLAS functions (isamax, icamax, isamin, icamin) may return incorrect values when the incx parameter is not equal to 1 on hardware not running Intel® AVX2 and later code paths. This issue will be fixed in the future oneMKL release.

System Requirements

Supported Architectures and Terminology - Intel® 64 Architecture refers to systems based on IA-32 architecture processors which have 64-bit architectural extensions (like Intel® Core™ architecture processors, running a 64-bit operating system such as Microsoft Windows* 10 x64 or a Linux* "x86_64" variant).

Note: Update releases (e.g., 2026.1) automatically inherit all OS requirements from the base major release (2026.0). Only operating systems marked with (+) are newly added in that update release, and those marked with (–) are removed. Asterisks(**) indicates Deprecations. All others remain unchanged and are not duplicated in the table.

Hardware Requirements

Supported Operating Systems

NOTE: For users using the Intel® oneAPI DPC++ Compiler, please refer to the Intel® oneAPI DPC++ Compiler system requirements here.

Software Requirements

Support Deprecated

2026.0

• Support for Windows* 10, Windows* Server 2019, RHEL* 8, Fedora* 43, Debian* 11 & 12, Amazon* Linux 2023 on CPU/Target has been deprecated.
• Support for SLES 15 SP4 on GPU has been deprecated.

2026.2

• Support for SLES 15 SP5, Ubuntu 22.04 LTS on CPU/Target has been deprecated.

Support Removed 

2026.0

• Support for Fedora* 42, Rocky Linux 9 on CPU/Target has been removed.
• Support for Windows* Server 2019, RHEL* 8.10 & 9.5, Ubuntu* 25.04 on GPU has been removed.

2026.2

• Support for RHEL* 10.0, SLES 15 SP4 on GPU has been removed.

Notices and Disclaimers

Intel technologies may require enabled hardware, software or service activation.

No product or component can be absolutely secure.

Your costs and results may vary.

© Intel Corporation. Intel, the Intel logo, and other Intel marks are trademarks of Intel Corporation or its subsidiaries. Other names and brands may be claimed as the property of others.

No license (express or implied, by estoppel or otherwise) to any intellectual property rights is granted by this document.

The products described may contain design defects or errors known as errata which may cause the product to deviate from published specifications. Current characterized errata are available on request.

Intel disclaims all express and implied warranties, including without limitation, the implied warranties of merchantability, fitness for a particular purpose, and non-infringement, as well as any warranty arising from course of performance, course of dealing, or usage in trade.