Average power analysis of sequential circuits using an autoregressive model

We present a new statistical technique for average power estimation in sequential circuits. Because of the feedback loops, power dissipations of sequential circuits in consecutive clock cycles are temporally correlated. The existence of data correlation makes it unsuitable to apply conventional techniques to average power inference, because the sample variance is no longer a maximum likelihood estimator. The convergence criterion derived from the biased variance estimation will be overly optimistic, causing power simulation to stop prematurely at a lower-than-specified estimation accuracy. To overcome this problem, we propose a systematic approach for modeling the power dissipation behavior of sequential circuits as an autoregressive random process. An accurate process variance can be obtained by the model parameters, which enables the derivation of a robust confidence interval of the average power. The interval is checked for convergence against a user-specified accuracy criterion. An iterative procedure is developed to invoke these steps repeatedly until the convergence specification is met. For a set of benchmark sequential circuits, this technique yields high accuracy and efficiency.