high performance computing on graphics processing units: hgpu.org

This paper presents an overview of the OpenCL 1.1 standard [Khronos 2012]. We first motivate the need for GPGPU computing and then discuss the various concepts and technological background necessary to understand the programming model. We use concurrent matrix multiplication as a framework for explaining various performance characteristics of compiling and running OpenCL code, and contrast this to native code on more traditional general purpose CPUs.