Abstract

This work presents a formal probabilistic approach for solving optimization problems in design automation. Prediction accuracy is very low especially at high levels of design flow. This can be attributed mainly to unawareness of low level layout information and variability in fabrication process. Hence a traditional deterministic design automation approach where each cost function is represented as a fixed value becomes obsolete. A new approach is gaining attention [15, 5, 2, 4, 12] in which the cost functions are represented as probability distributions and the optimization criteria is probabilistic, too. This design optimization philosophy is demonstrated through the classic buffer insertion problem [13]. Formally, we capture wirelengths as probability distributions (as compared to the traditional approach which considers wirelength as fixed values) and present several strategies for optimizing the probabilistic criteria. During the course of this work many problems are proved to be NP-Complete. Comparisons are made with the Van-Ginneken "optimal under fixed wire-length" algorithm. Results show that the Van-Ginneken approach generated delay distributions at the root of the fanout wiring tree which had large probability (0.91 in the worst case and 0.55 on average) of violating the delay constraint. Our algorithms could achieve 100% probability of satisfying the delay constraint with similar buffer penalty. Although this work considers wirelength prediction inaccuracies, our probabilistic strategy could be extended trivially to consider fabrication variability in wire parasitics.

Citations