A Map Reduce Framework for Programming Graphics Processors

Recent developments in programmable, highly parallel Graph- ics Processing Units (GPUs) have enabled high performance general purpose computation. We describe a framework de- signed for high performance GPU programming, built on Nvidia's Compute Unied Device Architecture (CUDA) plat- form. The framework is built around the Map Reduce ab- straction, which allows application developers to focus on their application, while enabling high performance GPU im- plementation. We show the utility of our framework by implementing Support Vector Machine training as well as classication, achieving speedups of up to 32 and 150 re- spectively over commonly used SVM software running on a CPU.