Dominik Göddeke (TU Dortmund)
GPU cluster computing for finite element
applications
This presentation explores GPU acceleration of parallel multigrid
solvers for Finite Element simulations on commodity based clusters.
The integration, based on a mixed precision iterative refinement
scheme, is minimally invasive in the sense that it requires no changes
to application code: Hardware acceleration is completely encapsulated
from the application level. We demonstrate good scalability and accurate
results for the prototypical Poisson problem and applications in
linearised elasticity and incompressible Navier-Stokes simulations.
Finally, we derive a performance model to accurately estimate the
achievable speed-up.