Section: New Results

High performance computing

High order DGTD-${\mathbb{P}}_p$ method on hybrid CPU/GPU parallel systems

Participants : Tristan Cabel, Stéphane Lanteri.

Modern massively parallel computing platforms most often take the form of hybrid shared memory/distributed memory heterogeneous systems combining multi-core processing units with accelerator cards. In particular, graphical processing units (GPU) are increasingly adopted in these systems because they offer the potential for a very high floating point performance at a low purchase cost. DG methods are particularly appealing for exploiting the processing capabilities of a GPU because they involve local linear algebra operations (mainly matrix/matrix products) on relatively dense matrices whose size is directly related to the approximation order of the physical quantities within each mesh element. We have initiated this year a technological development project aiming at the adaptation to hybrid CPU/GPU parallel systems of a high order DGTD-${\mathbb{P}}_p$ method for the numerical solution of the 3D Maxwell equations.