high performance computing on graphics processing units: hgpu.org

Due to their potential computation power, GPUs are often used for high performance computing. However, discrete GPUs are connected to the CPU via the PCIe bus, which can cause bottlenecks due to high latency and low bandwidth to the CPU. Lately, integrated GPUs have become more common, and due to being integrated on the CPU-chip, the bottleneck of the PCIe bus is reduced. Integrated GPUs are however less powerful than discrete GPUs, and as they share resources such as power and memory bandwidth with the CPU, heavy CPU utilization can cause a loss in performance. The aim of this thesis is to investigate the potential of integrated GPUs in high performance computing. This is done by comparing an integrated and a discrete GPU in various settings to see the impacts of the differences in latency and bandwidth, as well as the effects of the integrated GPU sharing resources such as power and memory bandwidth with the CPU. The results show that the less powerful integrated GPU outperforms the discrete GPU for smaller problem sizes due to the lower latency, even in situations when the shared resources are utilized by the CPU. Integrated GPUs can therefore be very useful when performing high performance computations on smaller datasets. The discrete GPU is however in many cases faster for larger datasets, especially for more arithmetically intense algorithms, due to a larger number of computation cores and dedicated memory.