Power macro-modeling using an iterative LS-SVM method

Abstract

We investigate the use of support vector machines (SVMs) to determine simpler and better fit power macromodels of functional units for high-level power estimation. The basic approach is first to obtain the power consumption of the module for a large number of points in the input signal space. Least-Squares SVMs are then used to compute the best model to fit this set of points. We have performed extensive experiments in order to determine the best parameters for the kernels. We propose a new method for power macromodeling of functional units for high-level power estimation based on Least-Squares Support Vector Machines (LS-SVM). Our method improves the already good modeling capabilities of the basic LS-SVM method in two ways. First, a modified norm is used that is able to take into account the weight of each input for global power consumption in the computation of the kernels. Second, an iterative method is proposed where new data-points are selectively added as support-vectors to increase the generalization of the model. The macromodels obtained compare favorably with those obtained using industry standard table models, providing not only excellent accuracy on average (close to 1% error), but more importantly, thanks to our proposed modified kernels, we were able to reduce the maximum error to values close to 10%.