A power macromodeling technique based on power sensitivity

Abstract

We propose a novel power macromodeling technique for high level power estimation based on power sensitivity. Power sensitivity defines the change in average power due to changes in the input signal specification. The contribution of this work is that we can use only a few points to construct a complicated power surface in the specification-space. With such a power surface, we can easily obtain the power dissipation under any distribution of primary inputs. The advantages of our technique are two-fold. First, the required parameters corresponding to each representative point can be efficiently obtained by only one symbolic power estimation run or by only one Monte Carlo based statistical power estimation process. This stems from the fact that power sensitivity can be obtained as a by-product of probabilistic or statistical power estimation runs. Second, the memory requirements for the macromodel are reduced to O(dn), where n is the number of primary inputs of a circuit and d is the number of representative points (d can be as small as 1 in some cases). Results on a number of benchmark circuits demonstrate the effectiveness of our technique.