Improving FDTD algorithm performance using block floating-point

Abstract

FDTD (finite-difference time-domain) algorithm is a memory intensive problem which requires more memory access in comparison to computing intensive problems. Existing FDTD solutions use either floating-point or fixed-point arithmetic. Floating-point solutions are less memory intensive but require more computation. Fixed-point solutions are the opposite — simpler computation but use up more memory bandwidth for the same precision. The method chosen in this paper includes using block floating-point arithmetic which is the middle ground. This approach requires less computation than floating-point arithmetic while at the same time using less memory access than plain fixed-point. The FDTD algorithm is converted from floating to fixed-point and then to block floating-point. The precision and memory access performance is compared.