A New Parallel Frequency-Domain Finite-Difference Algorithm Using Multi-GPU

Abstract

This letter presents a parallel frequency-domain finite-difference (FDFD) algorithm based on multi-graphic processing unit (GPU) applied to electromagnetic scattering computations to enhance the computational efficiency of the algorithm. The proposed algorithm parallelizes the solution of large-scale sparse matrices, distributing threads to the matrix-vector and vector-vector multiplication operations within decomposed sub-matrices to reduce the computational time. Moreover, we configure the OpenMP to optimize communication transfer between multiple GPUs, thereby improving computational efficiency. The simulation results show that compared with the conventional FDFD method, the proposed algorithm can enhance computational efficiency while ensuring accuracy.