Finite Element Analysis using Low-precision Computations

Abstract

While double-precision real numbers have been used for CAE simulations based on the finite element method, low-precision real numbers are used for deep learning for achieving much faster learning and inference. Due to overwhelming demand for deep learning, low-precision arithmetic is beginning to be introduced into general-purpose processors as accelerators. By using low-precision real arithmetic for finite element analysis, we can effectively use the arithmetic unit for deep learning and get advantages in terms of calculation speed and power consumption. In this study, in order to investigate the applicability of low-precision computation to finite element analysis, low-precision computation is firstly applied to the element integration process, and its effects are quantitatively evaluated.