We report on our experiences with integrating GPUs as fast, parallel
floating-point co-processors into the parallel FE package FEAST.
Since a full re-implementation of such a package is not feasible, we
identify the smoothing of an outer domain-decomposition multigrid
solver as a natural entry-point for a minimally invasive integration
of GPUs. We address the issue of limited computational precision
with a mixed precision iterative refinement approach and present
preliminary timing results on a commodity cluster enhanced with
GPUs.