high performance computing on graphics processing units: hgpu.org

Special hardware accelerators like FPGAs and GPUs are commonly introduced into a computing system as a separate device. Consequently, the accelerator and the host system do not share a common memory. Sourcing out the data to the additional hardware thus introduces a communication penalty. Based on a combination of a program’s source code and execution profiling we perform an analysis which evaluates the arithmetic intensity as a cost function to identify those parts most reasonable to source out to the accelerating hardware. The basic principles of this analysis are introduced and tested with a sample application. Its concrete results are discussed and evaluated based on the performance of a FPGA-based and a GPU-based implementation.