Abstract
As energy efficiency has become a primary concern, mobile and embedded system designers strongly desire for an efficient and highly accurate power estimation method. Since memory is an increasingly dominant power consumer, we particularly reexamine existing memory power models which are short of being satisfactory and propose a very efficient microcomponent-based approach with data-aware refinement for accurate system-level power estimations. The key of our approach is that we precisely identify the microcomponents activated by internal memory commands and accurately pre-calibrate the power consumption pattern of each microcomponent. To achieve very high accuracy, we further consider data effect by leveraging the fact that memory circuit is mainly doing data passing and both the dynamic and static power value can be refined in a linear fashion. With the runtime command sequence and timing information, we look up the command-activated microcomponents and their corresponding power patterns along with the runtime data-aware refinement. Our approach then generates very accurate and fast power analysis results. Our experiments show that the proposed approach produce accurate results of only 1.96% error rate in average.