Parameter tuning in SVM-based power macro-modeling

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. Based on this analysis, we propose an iterative method of improving the model by selectively adding new support vectors and increasing the sharpness of the model. The macromodels obtained confirm the excellent modelling capabilities of the proposed kernel-based method, providing both excellent accuracy on maximum error (close to 17%) and average (2% error), which represents an improvement over the state-of-the-art. Furthermore, we present an analysis of the dynamic range of power consumption for the benchmarks circuits, which serves to confirm that the model is able to accommodate circuits exhibiting a more skewed power distribution.