Model Variable Reduction Technique for High-Level Energy Estimation With An Accuracy Constraint

Energy consumption of a device is commonly modeled as a finite state machine. Finite state machines are easy to understand but a large number of states are required to model precise device power states. Therefore, system-wide energy estimation often uses highly simplified energy state machines because even an individual device requires a large number of states and transitions, but with a significant sacrifice of the estimation accuracy. In this paper, we introduce a systematic variable reduction technique applicable to a finite state machine for energy characterization aiming at cost-effective energy estimation. We reduce variables such as power coefficients and occurrence counters associated with the states and transitions, under a correlation coefficient constraint of the estimated power to the original full-size state machine. Our method is completely different from conventional state machine reduction techniques in that we preserve the original state machine structure while eliminate and/or merge variables into other variables as far as the correlation coefficient remains within the given tolerance. We propose a heuristic algorithm which effectively minimizes the number of variables within the range of specified prediction accuracy. We demonstrate that the number of variables is reduced by 96% at cost of less than 4% RMS error when the suggested algorithm is applied to a full DDR SDRAM.