In this article we discuss the influence of modern Computer hardware technologies on FEM software tools. As _moving data_ is much slower than _processing data_, most PDE software is not capable of fully exploiting fast processor architectures. The main reason for this is the use of highly unstructured meshes resulting from adaptive refinement techniques. We present an approach which combines modern adaptivity and solver concepts while reaching high MFLOP/s rates at the same time.
At first we describe our hierarchically working adaptive discretization strategy. Then we present appropriate solvers, the so called ScaRC solvers as generalization of multigrid and domain decomposition techniques. Finally we give some numerical results which support the theoretical expectations.