On mixture density and maximum likelihood power estimation via expectation-maximization

A maximum-likelihood estimation procedure for computing the average power consumption of VLSI circuits is proposed. The method can handle data that has a mixture-density with multiple components unlike most of the previous approaches. An iterative computational procedure based on the expectation-maximization principle is also discussed. This can be used to estimate the parameters of an arbitrary (but finite) number of components of the probability distribution of the simulated power data. Experimental results for ISCAS ´85 benchmark circuits and a large industrial circuit are given in order to validate the efficiency and practicality of the algorithm. Comparisons show that the proposed method estimates the multiple components (even those with a low probability of occurrence) while the Monte Carlo estimate captures only the most probable component.