GridNetOpt: Fast Full-Chip EM-Aware IR Drop Constrained Power Grid Optimization via Deep Neural Networks
TimeWednesday, December 8th6:00pm - 7:00pm PST
LocationLevel 2 - Lobby
Event Type
Networking Reception
Work-in-Progress Poster
Virtual Programs
Presented In-Person
DescriptionThis paper presents a fast full-chip EM-aware IR drop constrained optimization framework, named GridNetOpt, for on-chip power grid networks via a data-driven and learning-based IR drop modeling. The new method can naturally take advantage of differential characteristics of DNN for fast sensitivity computation. First, a CGAN model is trained where the training data comes from a state-of-the-art EM-IR analysis tool. Second, conjugate gradient based optimization framework is employed. Computing from the CGAN model, sensitivity information is obtained trivially through its auto-differentiation process. Numerical results show that the proposed GridNetOpt leads to order of magnitude speedup over the adjoint network method.