Regression modeling for subset selection in rare-event statistical circuit simulation

Yield estimation for high replication circuits such as SRAMs and flip flops is becoming a challenging task because a rare event in a circuit cell may impact significantly the whole system yield. The standard Monte Carlo Simulation is not efficient in detecting rare events because of the large number of simulations needed. The statistical blockade has been proposed to decrease the number of Monte Carlo simulations needed. In this paper it is shown that regression modeling can be used for parameter subset selection. This decreases the complexity of the problem while maintaining almost the same accuracy as the standard statistical blockade.