Optical metrology is a none-destructive measurement technology supplying reliable geometric information about small structures. It can be applied to any nm-sized object interacting with light, in particular to semiconductor components. Currently topologically new designs for active semiconductor components are discussed, most remarkably 3D VNANDS (flash memory) and FinFETs (logic). These structures are inherent 3D structures, in contrast to currently employed designs which are basically staggered 2D layouts. An important part of any optical metrology software is a fast forward solver for treating parameter-dependent scattering problems on high performance computers. Today's high performance architecture is characterized by parallel execution units on different execution levels. Important architecture approaches have to be considered, vectorization on CPUs, vector processing on hardware accelerators and parallelization across compute nodes. We want to establish a very competitive software/hardware combination based on hardware accelerators, general purpose GPUs (GPGPU, e.g. NVIDIA, AMD) as well as Intel Many Integrated Core Architecture (Intel MIC). In particular, we analyze, implement and apply a FEM-based scattering solver which combines standard FEM and, this is the new aspect, spectral representations for weakly parameter dependent waveguiding parts. The spectral representation requires the solution of generalized eigenvalue problems. The repeated solution of weakly parameter-dependent eigenvalue problems calls for an iterative solution adaption. We propose an automatic restarted preconditioned nonlinear Döhler-CG method as target for the acceleration engine. The project includes • modeling tasks from nanooptics, • algorithmic improvements on the numerical linear algebra side specific to the applications, and • hardware oriented designs and implementations. The team therefore involves, in a cross-disciplinary manner, scientists from the division Mathematics for Life and Material Sciences and the division Parallel and Distributed Computing.