Optimizing Parallel Reduction In Cuda To Reach GPU Peak Performance

Abstract

GPUs are massively multi threaded many core chips that have hundreds of cores and thousands of concurrent threads swith high performance and memory bandwidth. Now days, GPUs have already been used to accelerate some numerically intensive high performance computing applications in parallel not only used to graphic processing. This thesis aims primarily to demonstrate the programming model approaches that can be maximize the performance of GPUs. This is accomplished by a proof of maximum reach of bandwidth memory and get speed up from the GPU that used to process parallel computation. The programming environment that used is NVIDIA s CUDA, it is parallel architecture and model for generalpurpose computing on a GPU.