Tutorial Title:  Iterative solutions using programmable graphics processing units

Abstract

Graphics processing unit (GPU) development is rapidly advancing in speed and computational capabilities as it is driven by the desire for photorealism in real time desktop rendering. In comparison to Moore's law for CPU speeds, which predicts a doubling in processor speed every eighteen months, the GPU speed doubles every six months, effectively cubing Moore's law. This work investigates the feasibility of implementing an iterative algorithm on a programmable GPU (PGPU) using the Fuzzy C-Means (FCM) algorithm. The PGPU has been shown to provide significant reductions in computation times for a variety of non-iterative algorithms. However the feasibility of implementing complex iterative algorithms within a programmable graphics pipeline has yet to be determined. Our work shows that iterative algorithms can be implemented using PGPUs. Our PGPU FCM model was able to attain a speed up of of 1.7 to 2.1. This includes the time required to transfer data to and from the pipeline. This work was done in collaboration with Chris Harris, School of Computer Science and Software Engineering, University of Western Australia, Crawley, WA.