---
id: cupy-cuda12x
version: "14.1.1"
license: MIT
license_treatment: permissive
maintenance: active
---

# cupy-cuda12x — CuPy: NumPy & SciPy for GPU

License: permissive · Maintenance: active · Downloads: 3.2M/mo

## What it is and what it does

CuPy is a GPU-accelerated array library that mirrors NumPy and SciPy's API, allowing you to write numerical code that runs on NVIDIA GPUs instead of CPUs. It's built on CUDA 12.x and lets you leverage GPU parallelism for matrix operations, linear algebra, Fourier transforms, and other scientific computing tasks without rewriting your code from scratch. The package is actively maintained, supports modern Python versions (3.10–3.14), and is available as precompiled wheels for Linux (x86_64 and aarch64) and Windows.

You can install it standalone if you already have CUDA Toolkit 12.x on your system, or use the [ctk] extra to have pip bundle the necessary CUDA runtime components. It depends on NumPy and cuda-pathfinder for runtime operation. The library is in production-stable status and widely used in scientific and machine-learning workflows where GPU acceleration is needed.

Use it for:

- Accelerate large matrix operations and linear algebra computations by running them on GPU instead of CPU.
- Port existing NumPy code to GPU with minimal changes, using CuPy's drop-in compatible API.
- Build GPU-accelerated machine learning pipelines that need fast numerical array operations.
- Perform large-scale scientific simulations and data transformations that benefit from GPU parallelism.
- Combine CuPy with deep learning frameworks to handle numerical preprocessing on the same GPU device.

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

CuPy provides a NumPy- and SciPy-compatible array library that runs computations on NVIDIA GPUs using CUDA 12.x, letting you write GPU-accelerated numerical code with a familiar NumPy-like API.

Yes, if you have CUDA 12.x hardware and want GPU-accelerated NumPy-like operations. The package is actively maintained, permissively licensed, and widely used in production. Install friction is moderate—you need CUDA Toolkit 12.x or must use the [ctk] extra—but that's a one-time setup cost. No known security vulnerabilities. Not worth installing if you don't have NVIDIA GPU hardware or are locked into a different CUDA version.

## Install

pip install cupy-cuda12x
uv add cupy-cuda12x
poetry add cupy-cuda12x

## Installing cupy-cuda12x

Before you install: Medium install friction: requires CUDA Toolkit 12.x to be installed locally, or you can use the [ctk] extra to bundle CUDA components from PyPI. The package is actively maintained with recent releases and supports Python 3.10–3.14 across Linux, Windows, and ARM platforms.

License in practice: MIT license is permissive; you can use, modify, and distribute this package freely in commercial and private projects with minimal restrictions.

Quickstart:

pip install cupy-cuda12x
# or with bundled CUDA:
# pip install cupy-cuda12x[ctk]

import cupy as cp
x_gpu = cp.array([1, 2, 3])
result = cp.sum(x_gpu)

CUDA Toolkit 12.x must be installed locally on your system, or you must use the [ctk] extra to install CUDA components alongside the package.

Verify before relying:

- Performance gains relative to NumPy on typical workloads and hardware configurations.
- Memory overhead or compatibility gotchas when mixing CuPy and NumPy arrays in the same codebase.
- Support status for older CUDA 12.x minor versions (e.g., 12.0 vs. 12.6).

## Package facts

- License: MIT (permissive)
- Python support: supports_current
- Install friction: medium
- Maintenance: active
- Downloads: 3.2M/month (top 5,000 on PyPI)
- Known vulnerabilities: none known

## Tags

gpu array computing, numpy gpu acceleration, cuda python arrays, gpu scientific computing, cupy cuda 12, gpu numpy alternative, parallel array operations, gpu-computing, cuda, numerical-arrays

[View on SkillFed](https://skillfed.io/packages/cupy-cuda12x) · [View on PyPI](https://pypi.org/project/cupy-cuda12x/)