GPU-Accelerated LU and GMRES Solvers for Electromagnetic Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current solvers for the Method of Moments (MoM) in computational electromagnetics, such as LU and GMRES, become impractical for large-sized problems, requiring faster and more accurate methods that can handle larger problem sizes without sacrificing accuracy.
Innovation Solution
Implementing LU and GMRES solvers on a Graphics Processing Unit (GPU) using the Crout variant and optimized GPU techniques, specifically leveraging parallel processing to accelerate the solution of electromagnetic systems, including LU Decomposition and GMRES methods, to expand their applicability beyond small and medium-sized problems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If LU and GMRES solvers are used for solving dense systems of linear equations, then accuracy is improved, but processing time increases significantly for large problem sizes
Solution Approach 1:
The patent segments the dense N×N matrix into multiple blocks or sub-matrices, allowing parallel processing of different matrix blocks across multiple GPU cores. This segmentation enables LU and GMRES solvers to handle large-scale problems by distributing computational workload while maintaining solution accuracy through coordinated processing of segmented matrix operations.
Solution Approach 2:
The patent transitions from traditional sequential CPU-based solving to parallel GPU-based processing, adding a dimension of parallel computation. By leveraging the massively parallel architecture of GPUs, the solver can perform multiple matrix operations simultaneously, dramatically reducing processing time while maintaining the accuracy of LU and GMRES methods for large electromagnetic problems.
2Productivity
If accelerated methods are used to reduce processing time, then speed is improved, but solution accuracy deteriorates
Solution Approach 1:
The patent introduces the GPU as an intermediary computational platform that bridges the gap between CPU-based accuracy and raw processing speed. The GPU's parallel architecture acts as a mediator, enabling traditional accurate solvers (LU/GMRES) to run at accelerated speeds by distributing matrix operations across thousands of parallel cores without sacrificing numerical precision.
Solution Approach 2:
The patent changes the computational parameter from sequential processing to parallel processing by implementing matrix operations on GPU hardware. This parameter change transforms the execution model of LU and GMRES solvers, maintaining mathematical accuracy while achieving speedup through parallel computation of matrix factorizations and iterative updates.
3Adaptability or versatility
If problem size is increased to handle larger electromagnetic systems, then applicability is improved, but computational complexity increases making traditional solvers impractical
Solution Approach 1:
The patent segments large-scale electromagnetic problems into manageable matrix blocks that can be processed in parallel on GPU architectures. By dividing the N×N system matrix into smaller sub-matrices or blocks, the solver can handle problems with thousands of unknowns while keeping individual computational tasks within feasible complexity limits for each processing unit.
Solution Approach 2:
The patent adds parallel processing dimension to handle increased problem size. By leveraging GPU's thousands of parallel cores, the solver can process large electromagnetic systems with N×N matrices where N represents thousands of patches, transforming an otherwise intractable sequential problem into a manageable parallel computation.
Data Source
AI summary
Methods for increasing the processing speed of computational electromagnetic methods, such as the Method of Moments (MoM), may involve using efficient mapping of algorithms onto a Graphics Processing Unit (GPU) architecture. Various methods may provide speed/complexity improvements to either or both of: (1) direct solution via Lower-Upper (LU) factorization; and (2) iterative methods (e.g., Generalized Minimal Residual (GMRES) method).