Installation

Requirements

Supported GPUs

Numba supports all NVIDIA GPUs that are supported by the CUDA Toolkit it uses. Presently for CUDA 12 this ranges from Compute Capabilities 5.0 to 12.1 depending on the exact installed version, and for CUDA 13 this ranges from 7.5 to 12.1 (the latest as of CUDA 13.0).

Supported CUDA Toolkits

Numba-CUDA aims to support all minor versions of the two most recent CUDA Toolkit releases. Presently 12 and 13 are supported.

For further information about version compatibility between toolkit and driver versions, refer to CUDA Minor Version Compatibility.

Installation with a Python package manager

Conda users can install the CUDA Toolkit into a conda environment:

$ conda install -c conda-forge numba-cuda "cuda-version=12"

Or for CUDA 13:

$ conda install -c conda-forge numba-cuda "cuda-version=13"

Alternatively, you can install all CUDA 12 dependencies from PyPI via pip:

$ pip install numba-cuda[cu12]

CUDA 13 dependencies can be installed via pip with:

$ pip install numba-cuda[cu13]

If you are not using Conda/pip or if you want to use a different version of CUDA toolkit, CUDA Driver and Toolkit search paths describes how Numba searches for a CUDA toolkit.

Configuration

CUDA Bindings

Numba-CUDA uses the NVIDIA CUDA Python bindings for interacting with the CUDA Driver API. Numba-CUDA previously provided its own internal ctypes-based bindings; the public APIs exposing those bindings are kept for compatibility, but if you need to interact directly with the CUDA Driver or other CUDA libraries we recommend using the cuda-python package directly.

CUDA Driver and Toolkit search paths

Default behavior

Searches for the CUDA driver and toolkit libraries use the NVIDIA bindings’ built-in path-finding logic.

In addition to the CUDA toolkit libraries, which can be installed by conda into an environment or installed system-wide by the CUDA SDK installer, the CUDA target in Numba also requires an up-to-date NVIDIA driver. Updated NVIDIA drivers are also installed by the CUDA SDK installer, so there is no need to do both. When using the NVIDIA bindings (default), driver discovery is handled by their path-finding logic, and no Numba-specific environment variables are needed to locate libcuda.