numpy-mkl

This repository provides binary wheels for NumPy and SciPy, linked to Intel's high-performance oneAPI Math Kernel Library for Intel CPUs. The wheels are accessible through a custom Python Package Index (PyPI) and can be installed with pip or uv.

Installation

MKL-accelerated wheels are available for 64-bit versions of Linux and Windows. If using one of the recommended package manager below, there are no other prerequisites; all dependencies — including Python if using uv or Nix — are automatically installed by the package manager.

uv

# Create a Python project with uv.
uv init
uv add numpy scipy --index https://urob.github.io/numpy-mkl

pip

# Install globally or into active venv.
pip install numpy scipy --extra-index-url https://urob.github.io/numpy-mkl

nix

# Create and run virtual devshell using Nix.
nix flake init --template github:urob/numpy-mkl
nix develop .

manual

Wheels for manual installs can be downloaded here. Manual installs must install compatible versions of mkl-service, mkl, and its indirect dependencies. It is recommended to install mkl-service from this repository, which has been patched to detect and autoload the mkl library at runtime.

Cross-platform collaborations

MKL is only available on x86-64 architectures, excluding macOS systems. When using uv, one can use platform markers to automatically install MKL-linked versions of NumPy and SciPy when on a compatible system, and otherwise fall back to the default versions from PyPI.

Below is a simple illustration, which falls back to the PyPI packages on macOS.¹ To install the environment, copy the following snippet into pyproject.toml and then run uv sync. This will install a virtual environment in .venv, which can be activated on the command line or via most Python editors.

[project]
name = "example-project"
version = "0.1.0"
requires-python = ">=3.13"
dependencies = [
    "numpy>=2.2.6",
    "scipy>=1.15.2",
    "mkl-service>=2.4.2; sys_platform != 'darwin'",
]

[tool.uv.sources]
numpy = [{ index = "numpy-mkl", marker = "sys_platform != 'darwin'" }]
scipy = [{ index = "numpy-mkl", marker = "sys_platform != 'darwin'" }]
mkl-service = [{ index = "numpy-mkl", marker = "sys_platform != 'darwin'" }]

[[tool.uv.index]]
name = "numpy-mkl"
url = "https://urob.github.io/numpy-mkl"
explicit = true

Alternatives

The usual way to obtain MKL-accelerated NumPy and SciPy packages is through Anaconda or Conda-forge. The purpose of this repository is to provide an alternative for users who prefer to use pip or uv for package management. Other alternatives are listed below.

	Windows	Linux	Notes
This repository	Yes	Yes
Intel(r) Distribution for Python	Yes	Yes	Does not support NumPy 2.x
Numpy-mkl-wheels	Yes	No	Manual install only

Technical details

Linux wheels are built with gcc on Ubuntu 22.04. Windows wheels are built with msvc (numpy) and mingw-w64 (scipy) on Windows Server 2022. These compilers showed the most consistent runtime performance in a series of benchmarks, even in comparison to icx-compiled wheels.

All binaries are dynamically linked against the MKL library. The packages are patched to detect the library at runtime as long as mkl is installed anywhere accessible by Python. (This is indeed the case when using one of the recommended install methods above, which will automatically install mkl alongside NumPy or SciPy.)

References

• Intel(r) oneMKL Release Notes
• Intel(r) oneAPI Release Notes

Footnotes

1. More sophisticated checks can be added by combining with the platform_machine marker. In practices, distinguishing between macOS and other systems seems to be good enough for most use cases. ↩