CuPy

CuPy is an open-source array library for GPU-accelerated computing with Python.

  • AI
  • Tools & Utilities
CuPy is an open-source array library for GPU-accelerated computing with Python. CuPy utilizes CUDA Toolkit libraries including cuBLAS, cuRAND, cuSOLVER, cuSPARSE, cuFFT, cuDNN and NCCL to make full use of the GPU architecture.

Platform availability

Available on 0 of 10 platforms

  • Unraid (not listed)
  • TrueNAS (not listed)
  • Umbrel (not listed)
  • ZimaOS (not listed)
  • Proxmox (not listed)
  • Helm (not listed)
  • Mac (not listed)
  • Windows (not listed)
  • Linux (not listed)
  • Any server (not listed)

Health score?A 0-100 score computed nightly from four factors: maintained (40%: recent commits and steady activity), popular (25%: GitHub stars, log scale), easy to install (20%: platforms, install notes, setup guides, Docker image) and light to run (15%: minimum RAM, ARM support). When the source code has been checked, a fifth factor, security check (20%: automated scan for committed secrets, unsafe desktop-app settings, known-vulnerable dependencies and project hygiene; any serious finding caps it at 20), is added. Factors without data are left out and the rest are rescaled.

65/100 Good

Maintained
100last commit 18 days ago
Popular
8212,390 GitHub stars
Easy to install
250 install platforms, install notes, Docker image
Light to run
No data
Security check
152 serious findings, 6 minor, CI, tests, OpenSSF Scorecard 3.8/10

Based on 4 of 5 factors.

GitHub stars
12,390
Open issues
684
Last commit
2026-11-03
Latest release
v14.2.0
License
MIT
Activity
declining

Checked 18 days ago - source: GitHub (cupy/cupy)

Resources & Compatibility

ARM (e.g. Raspberry Pi) is not supported.

Checked 54 days ago - source: Docker Hub image tags and published docs

First-install notes

  • Install the CUDA-matched CuPy wheel Use the CuPy wheel for the CUDA (NVIDIA GPU platform) version your driver supports; the [ctk] extra can pull in needed CUDA components from PyPI if you don’t have a full CUDA Toolkit. If the driver or wheel version is wrong, CuPy will not start the GPU and GPU calls will fail. source
  • Add cuTENSOR/NCCL only if needed cuTENSOR and NCCL are optional GPU libraries for extra CuPy features, so install them with matching CUDA extras only if a recipe needs them. Without them, basic CuPy still works, but those specific functions are missing. source
  • Check the CUDA runtime actually in use After install, run `import cupy; print(cupy.cuda.get_local_runtime_version())` to see the CUDA runtime shared library actually present. This matters because CuPy’s default runtimeGetVersion reports the version CuPy was built with, not necessarily your machine’s CUDA. source

Fetched 53 days ago - each note links to its source

Alternatives (7)

  • Calcpad - Calcpad is a lightweight mathematical platform for cloud and desktop.
  • Julia - Julia is a powerful tool used to write computer programmes.
  • MATLAB - MATLAB is a numerical computing platform for data analysis, algorithm development, mathematical modelling, and visualisation, available as…
  • R (programming language) - R is a free software environment for statistical computing and graphics.
  • SageMath - SageMath is open-source mathematical software that combines many specialised tools into a single, easy-to-use system.
  • fxSolver - Online math solver, equation editor and database, graphing calculator and science/engineering problem helper.
  • wxMaxima - wxMaxima is a document based interface for the computer algebra system Maxima.

Checked 54 days ago - source: AlternativeTo

Common questions

How can I verify the CUDA runtime actually used by an installed CuPy build?

Import CuPy and print cupy.cuda.get_local_runtime_version() to see the CUDA runtime shared library actually present. The default version call may report the version CuPy was built with, not the version present on your machine.

Do I need to install cuTENSOR or NCCL for basic CuPy to work?

No, cuTENSOR and NCCL are optional GPU libraries for extra CuPy features. Install them with matching CUDA extras only if a recipe needs them; without them, basic CuPy still works but those specific functions are missing.

Where should I report a CuPy problem that comes from conda-forge?

Report it to cupy-feedstock. The maintainers can investigate whether it is a packaging issue in the conda-forge recipe or a real issue in CuPy.

Why might the first CuPy run be slower?

The first run can be slowed by CUDA kernel caching. The referenced thread asks for ways to reduce that initial slowdown while staying in high-level Python.

Answers sourced from docs.cupy.dev, github.com, github.com, www.reddit.com

Community

Collected 48 days ago - public community threads