Using Diffusion Model to Generate Design Solutions in very large-scale integrated computer-aided design (VLSI CAD)

Abstract

In the routing stage of back-end design in very large-scale integrated computer-aided design (VLSI CAD), congestion is used to evaluate routing performance. Congestion happens when the routing demand exceeds the routing resources available in a region. A route-friendly design is usually obtained through a routability estimator that provides congestion feedback to guide the kernel placer minimizing for placement objectives. We propose using a diffusion model to capture designs’ feature distributions to generate possible variations that might help further optimize placements for routability. In this work, we select three features as input to feed into a simple diffusion probabilistic model to synthesize routing feature design solutions. In our experiment, the visualization of the output that resembles that of the input design features was achieved.