Improved Model for Wire-Length Estimation in Stochastic Wiring Distribution

TL;DR

This paper introduces improved stochastic wiring distribution models that significantly enhance the accuracy of on-chip wire length estimation, addressing limitations related to Rent's exponent and proposing a new threshold for better approximation.

Contribution

The paper presents novel models with a new Rent's constant threshold that reduce estimation errors by up to 75%, improving upon existing methods.

Findings

28-50% reduction in estimation error

Impact of Rent's exponent analyzed

38-75% error reduction with new threshold

Abstract

This paper presents a pair of improved stochastic wiring distribution model for better estimation of on-chip wire lengths. The proposed models provide 28 - 50% reduction in error when estimating the average on-chip wire length compared to the estimation using the existing models. The impact of Rent's exponent on the average wire length estimation is also investigated to demonstrate limitations of the approximations used in some of the current models. To improve the approximations of the model a new threshold for Rent's constant is recommended. Simulation results demonstrate that proposed models with the new threshold reduce the error of estimation by 38 to 75 percent compared to the previous works.