cupy/cupy

Health v2 · maintenance only

Adoption not part of the score

12278 stars · 1132 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

CuPy is a NumPy/SciPy-compatible array library for GPU-accelerated computing with Python, running on NVIDIA CUDA or AMD ROCm. It acts as a drop-in replacement for NumPy/SciPy code and provides low-level access to CUDA features like custom kernels, streams, and runtime APIs.

Use cases

• accelerate numpy array computations on gpu
• run existing scipy code on gpu without rewriting
• write custom cuda kernels from python
• speed up numerical computing with gpu
• gpu-accelerated sparse matrix operations
• multi-gpu distributed array computations
• port numpy code to amd rocm gpus

When to choose

• you have existing NumPy/SciPy code that needs GPU acceleration with minimal changes
• you need low-level CUDA access (RawKernels, streams, runtime APIs) from Python
• you want a mature, actively maintained GPU array library with broad CUDA library support (cuBLAS, cuDNN, NCCL, cuTENSOR)

When to avoid

• you have no NVIDIA or AMD GPU available
• you need full NumPy/SciPy API coverage - some functions are missing or behave differently
• you need AMD ROCm support in production - it is still experimental
• you prefer a higher-level deep learning framework like PyTorch or JAX

Facets

library · maturity stable

machine-learning data-science math gpu-computing deep-learning machine-learning data-science gpu-computing deep-learning windows python numpy scipy cuda rocm gpu-acceleration array-library drop-in-replacement cudnn nccl algorithms linux gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="cupy/cupy")