cupy

CuPy is an open-source array library for Python that enables GPU-accelerated computing with NVIDIA CUDA. Designed to be a drop-in replacement for NumPy in many use cases, CuPy provides a cupy.ndarray class with an API that closely mirrors numpy.ndarray, including broadcasting, universal functions (ufuncs), and a wide range of linear algebra, random number generation, and advanced math operations.

CuPy organizes functionality into modules such as cupy.linalg for linear algebra, cupy.fft for fast Fourier transforms, cupy.random for random numbers, and cupy.sparse for sparse matrices. Additional features include elementwise kernels and reduction kernels for writing custom CUDA kernels, as well as support for CUDA streams and memory pools to manage asynchronous execution and minimize allocation overhead.

Interoperability with NumPy is straightforward; you can transfer data between CPU and GPU with cupy.asarray, cupy.asnumpy, or by using the DLPack interface for zero-copy sharing with other libraries. CuPy arrays live on the GPU memory, and data can be transferred back to the host when needed.

CuPy is typically installed via pip or conda, with wheels available for multiple CUDA versions. It requires an NVIDIA GPU with recent CUDA toolkit/drivers. It runs on supported platforms, mainly Linux and Windows, while CUDA is not available on macOS. CuPy serves as a backend for various higher‑level libraries and experimental frameworks, and is used in scientific computing, data analysis, and machine learning pipelines.

While CuPy provides broad compatibility with NumPy, some NumPy features are not yet implemented or differ in performance characteristics on the GPU. Users should be mindful of device memory limits and data transfer costs when designing algorithms.