Hierarchical Sequence Compaction for Power Estimation

Radu Marculescu, Diana Marculescu, Massoud Pedram

Abstract: This paper presents an effective technique forcompacting a large sequence of input vectors into a muchsmaller one such that when the two sequences are applied toany circuit, the resulting power dissipations are nearly thesame. Specifically, this paper introduces the hierarchicalmodeling of Markov chains as a flexible framework forcapturing not only complex spatiotemporal correlations, butalso dynamic changes in the characteristics of the inputsequence. The new framework has a high degree ofadaptability, i.e. the hierarchical model is dynamically grownaccording to the sequence behavior. Experimental resultsdemonstrate that large compaction ratios can be obtainedwithout a significant loss in accuracy (less than 5% on average)of power estimates.