Dresden 2014 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MA: Fachverband Magnetismus
MA 19: Poster I
MA 19.41: Poster
Tuesday, April 1, 2014, 13:00–15:30, P1
Multiple-GPU accelerated FEM micromagnetic simulations — •Attila Kákay1, Elmar Westphal2, and Riccardo Hertel3 — 1Forschungszentrum Jülich, Peter Grünberg Insitut (PGI-6), Deutschland — 2Forschungszentrum Jülich, PGI/JCNS-TA , Deutschland — 3Institut de Physique et Chimie des Matériaux de Strasbourg, Université de Strasbourg, CNRS UMR 7504, Strasbourg, France
The micromagnetic study of large and realistic systems, like rolled-up ferromagnetic nanotubes, artificial spin-ice lattices, or magnonic crystals poses a challenge to simulation studies, as their calculation may involve tens of millions of degrees of freedom. We recently made an important step toward the simulation of such large systems by adapting our micromagnetic Finite-Element software TetraMag to the massively parallel architecture of Graphical Processing Units (GPUs)[1]. But for large-scale simulations, the matrices required for the calculation of the effective fields can outgrow the memory capacity of a single GPU. By carefully redesigning our code, especially the magnetostatic field calculation and the integration of the Landau-Lifshitz-Gilbert equation, we can now distribute a simulation over several GPUs. This is achieved by reordering and splitting the matrices in a checkerboard style, which enables us to reduce and/or hide time-consuming data transfers that often have a large impact on the performance of multi-GPU algorithms. As a benchmark example we will discuss the spin wave dispersion and magnetic structures developing in the hysteresis loop of a 300 nm diameter and 4 µm long rolled-up Permalloy tube.
A. Kákay, E. Westphal, R. Hertel, IEEE Trans. Mag. 46, 2303 (2010)