GPU-accelerated mixed precision algebraic multigrid preconditioners for discrete elliptic field problems

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.