Title: LMM-IR: A Large-Scale Netlist-Aware Multimodal Framework for Static IR-Drop Prediction

Abstract:

Accurate and efficient static IR drop analysis is a cornerstone of integrated circuit design, yet it remains a bottleneck due to its computational intensity. The process can take several hours to complete, and resolving violations often requires multiple iterative cycles, significantly increasing the overall workload. Consequently, developing rapid and precise prediction methods is essential for streamlining the chip design workflow.

In this work, we introduce LMM-IR, a novel multimodal framework designed to accelerate this process. The system leverages a Large-Scale Netlist Transformer (LNT) to efficiently process SPICE files. Our primary innovation lies in the transformation of netlist topology into 3D point cloud representations, which allows for the scalable handling of complex circuits containing hundreds of thousands to millions of nodes.

LMM-IR integrates diverse data sources by encoding both netlist files and image data into a shared latent space. These encoded features serve as inputs for static voltage drop prediction, facilitating complementary insights from multiple modalities. Our experimental evaluations show that LMM-IR outperforms existing state-of-the-art algorithms, achieving the highest F1 score and the lowest Mean Absolute Error (MAE) compared to the leading solutions from the ICCAD 2023 contest.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC