New net models for spectral netlist partitioning

Abstract

Spectral approaches for partitioning netlists that use the eigenvectors of a matrix derived from a weighted graph model of the netlist (hypergraph) have been attracting considerable attention. There are several known ways in which a weighted graph could be derived from the netlist. However, the effectiveness of these alternate net models for netlist partitioning has remained unexplored. In this paper we first evaluate the relative performance of these approaches and establish that the quality of the partition is sensitive to the choice of the model. We also propose and investigate a number of new approaches for deriving a weighted graph model for a netlist. We show through test results on benchmark partitioning problems that one of the new models proposed here, performs consistently better than all the other models.