An Improved Objective for Cell Placement

To estimate the wiring area needed by the router to connect a signal net, most placement tools measure one half of the perimeter of the minimum rectangle enclosing all terminals of the net. In the past, this approach is reasonable because the half-perimeter value correlates well with the wiring area. As we are entering the deep-submicron era, the approach is no longer appropriate because the wiring delay must be characterized based on a distributed-RC model, in which not only the wiring area but also the wiring topology affects the wiring delay. In this paper, we show that the half-perimeter metric does not correlate well with the wiring delay under the distributed-RC model. We show that the radius of a net estimates the wiring delay more accurately than the half-perimeter metric does. We expand the acceptance criteria of a simulated annealing based placement tool to include moves that do not improve on the wiring length but do reduce the radius. Over all, for a set of benchmark circuits the critical path delays are improved up to 15%.