GPU Architecture Analysis And Power Modeling

In recent years, high performance computing architectures rapidly moved to the heterogeneous structure of CPU and GPU, and GPU plays an increasingly important role. But along with the significant increase in computing power and the architecture, the power consumption of GPU also gradually becomes a huge problem. How to use of GPU computing efficiently becomes a hot research issue. Recently NVIDIA Inc. developed a new architecture– Fermi, which is composed of many types of computing units and multi-level memory structure.The major contributions of my thesis include:We made proper abstraction of features of GPU architecture, especially about its computing units and multi-level memory. Based on tailored testing code in computing task, the throughput and latency of all instructions executed on computing units and the cost of all kinds of memory are measured. We also analyzed the features and the optimation of programming on GPU architecture.With sufficient understanding and parameters of GPU architecture, we build a power model of GPU architecture using tailored testing code. The model includes parameters regarding the energy cost of each instruction executed on the GPU architecture as well as accesses in different cache or memory. Using the test results and the profiler tools, we analyze even forecast the power of the applications. The accuracy of the power model is more than 90%.This research is sponsored partially by the National Hi-Tech 863 program under grant number 2009AA012201.