Maximizing GPU Utilization in High- Performance Sensor Signal Processing

Abstract

The purpose of this thesis work was to investigate the necessary steps to increase the utilization of a GPU and to eventually establish if the GPU is suitable as a processing unit in a challenging sensor signal processing chain. This was done by implementing the 2D-ESPRIT algorithm on a Quadro RTX 6000 GPU and analyzing the time that the GPU required to finish the necessary calculations. Then, the code was improved to increase computation speed and lower the execution time of the program. The first version of the implementation achieved a calculated GPU utilization of 0.82%. It was later found that the utilization of the GPU could be increased by dividing the data set into threads and CUDA streams which were computed concurrently. This increased the calculated utilization to 2.94%. Further, the main bottleneck of the 2D-ESPRIT algorithm was the singular value decomposition calculation, which if improved, would increase the utilization of the GPU further.