high performance computing on graphics processing units: hgpu.org

Supercomputers define the pinnacle of computational power and are an essential tool for solving vast scientific computational problems. They employ increasingly parallel architectures to ever increase their nominal peak performance and to allow them to solve larger problems. Employing the vast amount of computation power is however difficult and optimising for many-core architectures has become of paramount importance. This dissertation aims to obtain an in-depth understanding of modern high-performance accelerators by theoretical study and practical in-depth exploration of the hardware characteristics. The focus lies with the Intel Xeon Phi co-processor, which powers the fastest supercomputer to date. A generic and portable benchmark framework that caters to highly configurable and reusable (micro-) benchmarks has been developed in support of the practical exploration. The effective exploration resulted in the successful identification of new and interesting properties regarding the behaviour of the Xeon Phi.