Efficient ANN-based interconnect delay and crosstalk modeling

Abstract

The dominance of system performance by interconnect delay in deep-submicron design presents many challenges to physical design tool developers. This paper presents an efficient ANN-based technique for modeling interconnect crosstalk in integrated circuits. ANN models for user-defined interconnect primitives called wirecells are trained and tested using a database created using a suitable simulation package. For fixed wirecell length and geometry, inputs to the ANN include signal frequency, input voltage amplitude, near and far end termination impedances. Outputs derived from the ANN include crosstalk voltage peak and RMS values and spectral composition. Experimental results demonstrate the ability of this approach to successfully predict coupled noise in modest cpu times compared with existing approaches.