high performance computing on graphics processing units: hgpu.org

Optimizing program code is an essential process for High-Performance Computing and in general. Due to a trend in the last years of employing graphics cards as accelerators for systems and due to a universal gain of the importance of GPUs, optimizing GPU code is crucial in order to achieve the best possible performance of a GPU program. The optimization process works best with fundamental background knowledge about the GPU(s) and their structure on which the program will be executed. However, a lot of the relevant information especially regarding the overall memory structures is not or only vaguely documented. For this purpose, this thesis presents the design and implementation of a tool to capture the memory topology of GPUs. This work only considers the GPUs of Nvidia whereas AMD or other manufacturers are not covered.