Evaluating multi-GPU computing capabilities of Numba and CuPy

Abstract

In this paper, we evaluate the performance of Numba and CuPy in multi-GPU configurations, focusing on both strong and weak scalings. We employ two benchmark problems: pseudo-random number generation and one-dimensional Monte Carlo radiation transport in a purely absorbing medium. In the experiments, we compare Numba, CuPy, and CUDA C implementations under both default and optimized conditions implemented on the NVIDIA DGX-2 server platform. In the default setup, CUDA C delivers better performance and the highest energy efficiency. However, we demonstrate that CuPy can achieve substantial speedups in optimized mode, though this requires extensive code modifications. Numba shows competitive performance in cases with minimal data transfer to global memory, but its scalability and energy efficiency are limited by CPU-side random number generator state initialization bottlenecks, and it benefits less from optimizations compared to CuPy.