cupy

CuPy's API closely mirrors NumPy, allowing existing code like cp.arange to run on GPUs with minimal changes, as shown in the README example where array operations work identically.

Supports both NVIDIA CUDA and AMD ROCm architectures, offering flexibility across GPU hardware, though ROCm support is noted as experimental in the installation guide.

Provides direct access to CUDA features such as RawKernels and Streams, enabling advanced users to fine-tune performance for critical applications, as mentioned in the low-level access section.

Includes cuSignal functionality for accelerated signal processing workflows, reducing the need for separate libraries, with integration noted starting from v13.0.0.