Mixed- Precision Algorithm in Electromagnetic Finite Element Method

Single-precision algorithms are known to offer superior computational efficiency compared to double-precision algorithms. However, achieving the same level of computational accuracy as double-precision algorithms in single-precision arithmetic can be challenging. Mixed-precision algorithms have emer...

Full description

Saved in:
Bibliographic Details
Published in:2023 International Applied Computational Electromagnetics Society Symposium (ACES-China) pp. 1 - 3
Main Authors: Zhi-ming, Li, Hao-xiang, Wu, Sheng, Zuo, Xi, Chen
Format: Conference Proceeding
Language:English
Published: Applied Computational Electromagnetics Society (ACES) 15.08.2023
Subjects:
Approximation algorithms
Computational efficiency
Computational modeling
Electromagnetic Finite Element method
Finite element analysis
Iterative methods
Mathematical models
Memory management
mixed-precision algorithm
Sparse matrix direct solver
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Single-precision algorithms are known to offer superior computational efficiency compared to double-precision algorithms. However, achieving the same level of computational accuracy as double-precision algorithms in single-precision arithmetic can be challenging. Mixed-precision algorithms have emerged as a promising solution to this problem. In this study, we employ a mixed-precision algorithm to solve the matrix equation Ax=b generated by the electromagnetic finite element method. Specifically, single-precision arithmetic is utilized in the matrix factorization phase, whereas double-precision arithmetic is applied in the residual calculation and iterative refinement phase. Our experimental results demonstrate that the mixed-precision algorithm achieves computational efficiency that is approximately 1.5 times higher than that of the double-precision algorithm, and occupies 40% less memory. Additionally, the computational accuracy of the mixed-precision algorithm is found to meet the requirements of double-precision arithmetic.
DOI:10.23919/ACES-China60289.2023.10249847