CFIRSTNET: Comprehensive Features for Static IR Drop Estimation with Neural Network

TL;DR

CFIRSTNET addresses the challenge of fast, accurate static IR drop estimation by fusing image-based inputs with netlist-derived features via a CNN-based encoder-decoder. It introduces Hypothetical IR Drop Distillation and multi-feature maps (wire resistance and resistive distance) processed by a ConvNeXtV2-based encoder and FPN decoder to generate high-resolution IR drop maps. On the ICCAD CAD Contest 2023 open-source benchmark, CFIRSTNET achieves lower MAE and max errors, higher hotspot detection F1, and substantial runtime speedups compared with SPICE-based methods and prior ML approaches. The method is technology-agnostic and scalable to irregular PDNs and macros, enabling practical integration into CAD flows. The work demonstrates that combining both image- and netlist-based features with a tailored CNN yields state-of-the-art static IR drop predictions.

Abstract

IR drop estimation is now considered a first-order metric due to the concern about reliability and performance in modern electronic products. Since traditional solution involves lengthy iteration and simulation flow, how to achieve fast yet accurate estimation has become an essential demand. In this work, with the help of modern AI acceleration techniques, we propose a comprehensive solution to combine both the advantages of image-based and netlist-based features in neural network framework and obtain high-quality IR drop prediction very effectively in modern designs. A customized convolutional neural network (CNN) is developed to extract PDN features and make static IR drop estimations. Trained and evaluated with the open-source dataset, experiment results show that we have obtained the best quality in the benchmark on the problem of IR drop estimation in ICCAD CAD Contest 2023, proving the effectiveness of this important design topic.

Figures (7)

• Figure 1: Illustration of a typical PDN structure: a resistive network between the power sources (C4 bumps) and the standard cells.
• Figure 2: Extracted features and the CFIRSTNET prediction flow: Along with the provided image-based data, CFIRSTNET extracts augmented features from the netlist-based data and processed by the custom CNN.
• Figure 3: Metal stripe and circuit segments.
• Figure 4: The detailed illustration of custom CNN model architecture in Figure \ref{['fig: flow']}. CFIRSTNET extracts the latent representation of the PDN from the model input features in the encoder stage (purple), aggregates features of various resolutions with the FPN (blue), and generates a high-resolution IR drop estimation map with the custom reconstructor.
• Figure 5: Comparison between IR drop maps under different resolutions. A lower resolution (12$\mu$m x 12$\mu$m) can not represent the actual IR drop values.