high performance computing on graphics processing units: hgpu.org

Emergence of new Graphical Processors for general purpose computing presents new challenges for application developers. Graphical Processors vary in terms of number of processor cores per chip, processor speed and memory subsystems. NVIDIA’s CUDA provides a C-like abstraction layer for software developers to implement their applications on GPUs often with little knowledge of the underlying hardware and they are forced to work with high-level descriptions documented by the manufacturer. Substantial knowledge of the hardware architecture will be useful for harvesting the full potential of GPU architectures while trying to solve complex parallel programming problems. This work reports the measurements and characterization of performance of several NVIDIA GPU’s using micro benchmark analysis. Our thesis uses and adapts the CUDA Micro-benchmarks [8] and SHOC benchmarks [9] to characterize the important aspects of NVIDIA’s GTX200 series GPU- architecture machine (GTX280) and Fermi series – architecture machines (GTX580, Tesla C2050). The investigation is conducted by performing a micro architectural analysis of these machines and comparing their basic performance parameters. This thesis presents an experiment based methodology for characterizing the properties of the arithmetic pipelines. We also measure the global and shared memory latency and bandwidth of these machines and validate the hardware characteristics presented in CUDA programming guide. We hope that the insights from this work will be useful for improving the analysis and performance optimization of CUDA programs.