Simulation Research On Multi-Core Stream Architecture And Power Consumption Of GPGPU

With the progress of microprocessor technology, CMP has been the mainstream of design. The multi-core stream processor shows tremendous computing capability and it had advantages in area utilization, average power consumption and programming flexibility. As a typical multi-core processor, GPGPU has made a great impact on dealing with dense data and parallel computing. By studying the architecture of GPGPU, we can explore the direction of computer architecture development, which provides a way to produce homemade general-purpose stream processor.Simulator is an effective tool for researching processor architecture. The architecture of GPGPU has the characters in both multi-core processor and stream processor, which make it quite different from traditional processor architecture, and so it acquires new simulation technology and methodology. Therefor, we choose the GPGPU of NVIDIA Corp, which is used in academia widely, to do related research.This thesis analyses the evolvement and architectural specialty of GPGPU, and with the study in the programming model which is named CUDA and the multithreading executing mode, we detailedly discuss the main idea of multi-core stream processor. This thesis fully makes use of the technology and methodology of the existing simulator named GPGPU-Sim, and by extending the software and perfecting its function, we utilize the programming interface and algorithm of Watch, which is a famous power simulator, to establish the architectural power model of GPGPU. The simulation results show that the GPGPU simulator is able to verify the function and GPGPU reliably. When allocated more threads to execute, the GPGPU has the better speedup, for the stream multiprocessors are filled with the threads more efficiently. The number of multiprocessors is the main factor to determine the performance of GPGPU, and at the same time, the configuration of pipeline, the DRAM scheduler and the clock frequence can also affect the performance. Change the memory hierarchy or the programming mode may have great effect on the performance of GPGPU. For the application which has ordered data and unique flow of execution, it has the best performance to fully use coalescing mechanism without data cache; but for more general-perpose computing applications, it is better to utilize data cache. On the other hand, the power consumption of GPGPU increases when the number of multiprocessors or the number of threads increases, when the memory hierarchy and the programming mode are also important.