Title :
GPU-accelerated mixed precision algebraic multigrid preconditioners for discrete elliptic field problems
Author :
Richter, Chris ; Schöps, Sebastian ; Clemens, Markus
Author_Institution :
Electromagn. Theor., Bergische Univ. Wuppertal, Wuppertal, Germany
fDate :
March 31 2014-April 1 2014
Abstract :
In this paper the use of a mixed precision implementation of Krylov subspace methods with multigrid preconditioners is proposed for solving the large linear systems stemming from Finite Element or Finite Difference method discretizations of elliptic problems as they occur e.g. in electrostatics. The computational limits in speed and memory are discussed using the numerical example of a large scale 3D high voltage isolator model.
Keywords :
computational electromagnetics; finite difference methods; finite element analysis; graphics processing units; GPU; Krylov subspace method; discrete elliptic field problem; finite difference method discretization; finite element method; large scale 3D high voltage isolator model; linear system; mixed precision algebraic multigrid preconditioner; CUDA; Conjugate Gradients; Finite Elements; GPU Programming; Mixed Precision;
Conference_Titel :
Computation in Electromagnetics (CEM 2014), 9th IET International Conference on
Conference_Location :
London
Electronic_ISBN :
978-1-84919-817-2
DOI :
10.1049/cp.2014.0185
