high performance computing on graphics processing units: hgpu.org

General-Purpose computing using Graphics Processing Units (GPGPU) has been an area of active research for many years. During 2009 and 2010 much has happened in the GPGPU research field with the release of the Open Computing Language (OpenCL) programming framework and the new NVIDIA Fermi Graphics Processing Unit (GPU) architecture. This thesis explores the hardware architectures of three GPUs and how well they support general computations; the NVIDIA Geforce 8800 GTS (the G80 architecture) from 2006, the AMD Radeon 4870 (the RV700 architecture) from 2008 and the NVIDIA Geforce GTX 480 (the Fermi architecture) from 2010. Special concern is given to the new Fermi architecture and the GPGPU related improvements implemented in this architecture. The Lukas-Kanade algorithm for optical flow estimation has been implemented in OpenCL to evaluate the framework and the impact of several different parallel application optimizations. The RV700 architecture is not well suited for GPGPU. The performance of the G80 architecture is very good taking its relative age into account. However, much effort must be spent optimizing a parallel application for the G80 before full performance is obtained, a task that can be quite tedious. Fermi excels in all aspects of GPGPU programming. Fermi’s performance is much higher than that of the RV700 and the G80 architectures and its new memory hierarchy makes GPGPU programming easier than ever before. OpenCL is a stable and competent framework well suited for any GPGPU project that would benefit from the increased flexibility of software and hardware platform independence. However, if performance is more important than flexibility, NVIDIA’s Compute Unified Device Architecture (CUDA) or AMD’s ATI Stream might be better alternatives.