PDNNet: PDN-Aware GNN-CNN Heterogeneous Network for Dynamic IR Drop Prediction

TL;DR

The paper tackles dynamic IR drop prediction in power delivery networks by moving beyond CNN-based representations to a graph-based PDNGraph that encodes fine-grained PDN structure and cell-PDN relations. It introduces PDNNet, a dual-branch GNN-CNN network where the GNN branch learns from the PDNGraph and the CNN branch captures temporal IR drop dynamics, with a fusion module producing the final IR drop map. Experiments on CircuitNet show that PDNNet achieves superior accuracy compared with state-of-the-art CNN-based methods and delivers up to 545× speedups over commercial tools, demonstrating the practical impact for IC design closure. The work provides a new framework for PDN-aware deep learning in IC design, with interpretable components and strong generalization across designs.

Abstract

IR drop on the power delivery network (PDN) is closely related to PDN's configuration and cell current consumption. As the integrated circuit (IC) design is growing larger, dynamic IR drop simulation becomes computationally unaffordable and machine learning based IR drop prediction has been explored as a promising solution. Although CNN-based methods have been adapted to IR drop prediction task in several works, the shortcomings of overlooking PDN configuration is non-negligible. In this paper, we consider not only how to properly represent cell-PDN relation, but also how to model IR drop following its physical nature in the feature aggregation procedure. Thus, we propose a novel graph structure, PDNGraph, to unify the representations of the PDN structure and the fine-grained cell-PDN relation. We further propose a dual-branch heterogeneous network, PDNNet, incorporating two parallel GNN-CNN branches to favorably capture the above features during the learning process. Several key designs are presented to make the dynamic IR drop prediction highly effective and interpretable. We are the first work to apply graph structure to deep-learning based dynamic IR drop prediction method. Experiments show that PDNNet outperforms the state-of-the-art CNN-based methods and achieves 545x speedup compared to the commercial tool, which demonstrates the superiority of our method.

Figures (8)

• Figure 1: Example of IR drop maps with different PDNs under the same current load. The left part is PDN architecture. High metal (e.g., M8 and M7) and low metal (e.g., M1) are distinguished with different color. The corresponding IR drop map is in the right position. The color approaching Red denotes a higher IR drop value, and approaching Blue denotes lower.
• Figure 2: The illustration of PDN grid in IR drop analysis.
• Figure 3: Analysis of simulation-based method, conventional deep-learning prediction methods, our proposed PDNGraph, and PDNNet framework.
• Figure 4: Illustration of our PDNGraph construction. Figure better viewed in color mode. From left to right is (a) The whole graph on IC layout. (b) The fine-grained part inside PDN strip. (c) The fine-grained part outside PDN strip.
• Figure 5: The figure shows how the directions of edges between nodes in PDNGraph are determined. We categorize the relationship between the PDN location and the current node into four cases.