This diploma thesis has three topics: Make it possible to use the Fourier transform on CUDA devices using the NVIDIA CUFFT library, port a sub system of a motion estimator to the graphics card and finally the design and implementation of an abstraction layer for many core systems.

The problem of using the CUFFT library is its high memory consumption, especially for images of more than 8 Megapixel in size. Evaluating this shows, that it can bring current standard graphics cards to their limit. The solution to this problem is to split up the two-dimensional Fourier transform into two one-dimensional transforms. Its mathematical definition allows to transform larger images without any loss of performance.

The motion estimator sub system is the so called cleaning part. Its task is to smoothen the vectors which where created by the estimation when trying to find the direction a block of a frame moved to. It will be shown that todays CUDA hardware suffers from heavy performance losses when compared to a CPU implementation.

The third task was to design and implement an abstraction layer for the hardware of a many core system. Its particularly features are the hardware abstraction which allows to write code once and then compile and use it with every architecture available on the system. Additionally, it allows to split up its given tasks and distribute them over the computation devices available.