Boost NumPy Performance 10× with CuPy: GPU Acceleration Guide

NumPy, a core Python library for multi‑dimensional arrays, already speeds up many operations compared to pure Python, but its performance is limited to CPU cores.

CuPy mirrors the NumPy API but runs on NVIDIA GPUs via CUDA, offering much higher parallelism.

Installation is as simple as pip install cupy. After installing, import it as import cupy as cp just like NumPy.

Benchmark tests on a PC with an i7‑8700K CPU, 1080 Ti GPU, 32 GB DDR4 RAM and CUDA 9.0 show dramatic speedups. Creating a 1000×1000×1000 array of ones takes 1.68 s with NumPy on CPU but only 0.16 s with CuPy on GPU—a 10.5× improvement.

Multiplying the array by 5 further highlights the gap: NumPy needs 0.507 s while CuPy finishes in 0.00071 s, a 714× boost.

For a series of three operations (multiply by 5, element‑wise multiplication, addition to itself) on larger data, NumPy requires 1.49 s versus CuPy’s 0.092 s, a 16.2× gain.

Performance scales with data size; when the number of elements exceeds 10 million, GPU acceleration becomes pronounced, though for smaller arrays CPU may still be faster. Sufficient GPU memory is required for very large datasets.