GPU Acceleration of Algebraic Multigrid Preconditioners for Discrete Elliptic Field Problems

The simulation of coupled electromagnetic/thermal problems with high resolution requires efficient numerical schemes. High-performance computing languages like CUDA help in unlocking the massively parallel capabilities of graphic processor units (GPUs) to accelerate those calculations. This reduces the time needed to solve real-world problems. In this paper, the speedup is discussed, which is obtained using NVIDIA's recently presented Kepler architecture as well as by GPU-accelerated algebraic multigrid preconditioners. In particular, extended memory allows for the solving of larger problems with more degrees of freedom without swapping. We discuss a new host-based multigrid setup for GPU-accelerated iterative solvers.