Evaluation of Support Vector Machines for PCB based Power Delivery Network Classification

Abstract

In this paper the performance of support vector machines (SVMs) is investigated to classify printed circuit board (PCB) based power delivery networks (PDNs). For different decoupling capacitor (decap) distributions on a PCB the impedance of the PDN is evaluated. The target impedance (TI) is used as separation condition of the two classes meeting or violating the TI. It is shown that the preprocessing of the PCB parameters and decap distributions have a strong impact on the prediction accuracy of SVMs. Furthermore, variations of the PCB structure with respect to the geometry are investigated. It is shown using extended SVMs with geometry features as additional input feature that it is possible to achieve similar prediction accuracies in comparison with a collaboration of multiple SVMs. Finally the performance of power integrity (PI) classifications by SVMs in comparison with artificial neural networks (ANNs) is discussed.