Abstract:General purpose computation on graphics processing units (GPGPU) has opened up a host of possibilities for high performance computing on commodity hardware. We show how an interesting texture segmentation algorithm can achieve 35x- 50x speedups on the GPU. We also show that portions of the algorithm can even approach a 300x speedup. We also demonstrate that portions of the algorithm that form bottlenecks in the CPU implementation are not bottlenecks on the GPU. However, most importantly, we argue that most of the computational power of graphics processing units is still unharnessed and show how limitations in the architecture and programming model inhibit even faster GPGPU algorithms.

Paper: