Simple eigenvector-based circuit clustering can be effective [VLSI CAD]

Clustering has proven effective in improving the quality of VLSI netlist partitioning and placement algorithms. A wide variety of clustering schemes have been proposed, including random walks, iterative matching, and fairly complicated spectral techniques. We use eigenvectors to compute a clustering, but do so in the simplest, most obvious manner. Our algorithm first computes a d-digit code for each module v/sub i/ according to the signs of the i/sup th/ entries in a set of d eigenvectors. Then, modules with the same code are assigned to the same cluster. Despite its simplicity, this new clustering algorithm is strongly motivated by theoretical results for both spectral bipartitioning and multi-dimensional vector partitioning. The algorithm also has linear time complexity (not including the eigenvector computation) and is at least as effective as previous clustering algorithms in terms of two-phase Fiduccia-Mattheyses bipartitioning.