GPU-Accelerated Mixed-Precision FDFD Method for 3-D Electromagnetic Problems

In this paper, a mixed-precision parallel finite difference frequency-domain (FDFD) scheme is proposed for the efficient solution of 3D electromagnetic problems. In the proposed scheme, the mixed-precision technique is introduced into the BiConjugate Gradient (BiCG) solver: computation-intensive sparse matrix–vector multiplications (SpMV) are performed in single precision, while inner products and residual updates are carried out in double precision. Groupwise update technique is used to eliminate the impact of rounding errors. The entire process is fully parallelized on GPU, leveraging its optimization for lowprecision arithmetic. Experimental results show that, compared to existing GPU-based FDFD methods, the proposed method significantly reduces memory consumption and computational cost while maintaining solution accuracy.