A Statistical Learning Based Modeling Approach and Its Application in Leakage Library Characterization

In statistical analysis, modeling circuit performance for non-linear problems demands large computational effort. In semi-custom design, statistical leakage library characterization is a highly complex yet fundamental task. The log-linear model provides an unacceptable poor accuracy in modeling a large number of standard cells. To improve model quality, simply increasing model order is not practicable because it leads to an exponential increase in run time. Instead of assuming one model type for the entire library beforehand, we developed an approach generating a model for each cell individually. The key contribution is the use of a cross term matrix and an active sampling scheme, which significantly reduces model size and model generation time. The effectiveness of our approach is clearly shown by experiments on industrial standard cell libraries. As we regard the circuit block as a black box, our approach is suitable for modeling various circuit performances.