CSR-AES: Characterizing the Cooperative Parallel Behavior of GPUs and Multicore CPUs

点击翻译按钮获取中文摘要

Improving the performance of microprocessors by increasing the clock speed is no longer practical. In response, microprocessor manufacturers have chosen to increase performance through the use of multiprocessing. By populating a single die with multiple, less complicated processors, manufacturers are creating CPUs that provide higher computational performance per watt and per square millimeter. At the same time, GPU chips have used the same ideas to vastly improve their aggregate performance, although they address different forms of computational problems.This project is examining these two complementary resources as a single cooperative execution system. It is characterizing how to best take advantage of each, and ascertain the ?sweet spot? of assigning different computational functions to the appropriate processing units based on the nature and size of what is being computed. It will not just produce an optimal point solution for one particular problem, but will generalize the overall problem in terms of: the type and size of the data computation, the speed of the CPUs and GPUs, the number of CPU and GPU processors, and the bus speed.The result of this research will be to create a desktop ?mini-supercomputer? for solving compute-intensive scientific and rendering problems. Because of its multicore CPU/GPU mix, it will have more combined compute power than any of its desktop predecessors, but it will look to an application developer as a single compute environment. Best of all, it will have no more cost and footprint than the graphical desktop system that the user would have bought anyway.