Stefan Turek: [FW] Can we realize much faster FEM-based CFD solvers (to compete with AI)?

Speaker
Stefan Turek, Dortmund University of Technology
Abstract
We discuss how modern HPC facilities with massively parallel together with lower precision accelerator hardware might be exploited in numerical flow simulations based on FEM techniques so that a very high computational, numerical and hence energy efficiency can be obtained. We shortly illustrate how “parallel-in-space simultaneous-in-time” Krylov-Multigrid approaches in combination with global-in-time Newton methods and special „Pressure Schur Complement“ techniques can be designed which should allow a much higher degree of parallelism for the resulting sequences of time-dependent convection-diffusion velocity and pressure Poisson-type subproblems. To exploit current accelerator hardware in lower precision, mainly working in single or even half precision, we shortly explain the concept of “prehandling” for the discretized systems of equations. We provide numerical results as “proof-of-concept” and discuss open problems and challenges for incompressible flow problems in industrial applications including complex rheological behaviour.