Is probabilistic congestion estimation worthwhile?

  • Authors:
  • Jurjen Westra;Patrick Groeneveld

  • Affiliations:
  • Eindhoven University of Technology, Eindhoven, The Netherlands;Eindhoven University of Technology, Eindhoven, The Netherlands

  • Venue:
  • Proceedings of the 2005 international workshop on System level interconnect prediction
  • Year:
  • 2005

Quantified Score

Hi-index 0.00

Visualization

Abstract

Fast congestion prediction is essential for congestion reduction techniques at different stages of the flow. Probabilistic congestion estimation methods model congestion after placement by considering the probability a wire will be routed over different areas of the routing region. They are popular because they are much faster than traditional global routing.In this paper, two congestion estimation tools are presented. The first one is an implementation of a probabilistic method called pce and appears to be very fast in comparison with comparable methods. The second one called FaDGloR is new and based on global routing techniques. Surprisingly, FaDGloR is about as fast as pce. The reason is that it is tuned towards congestion estimation and speed contrary to global routers that are tuned towards wire length reduction and routing as many wires as possible. Special focus is on congested areas of the chip, the areas that may prevent a design to be routable, and FaDGloR more accurately predicts these areas than pce. Both tools are tested on designs varying from impossible to route to easily routable. Previous papers focussed on accurately modeling congestion of the final, routable design. In practice, also unroutable designs need to be evaluated. The results presented in this paper indicate that global routing based methods are probably more worthwhile than probabilistic methods.