Si-GT

This repository provides the official implementation of the ICLR 2026 poster "Si-GT: Fast Interconnect Signal Integrity Analysis for Integrated Circuit Design via Graph Transformers".

Si-GT is a graph Transformer model designed for fast interconnect signal integrity analysis in integrated circuit design. It supports two key tasks: crosstalk delay prediction and crosstalk glitch prediction. For crosstalk delay, Si-GT predicts the net delay on both aggressor and victim nets. For crosstalk glitch, Si-GT predicts the peak noise voltage and noise width.

Model Overview

Si-GT elaborates three key designs: (1) virtual NET token to encode net-specific signal characteristics and serve as net-wise representation, (2) mesh pattern encoding to embed high-order mesh structures at each node while distinguishing uncoupled wire segments, and (3) intra-inter net (IIN) attention mechanism to capture structures of signal propagation path and coupling connections.

Dataset Download

The datasets can be downloaded using gdown.

pip install gdown

# Crosstalk delay dataset
gdown 1Ur_CP4HmWD1J5RhOrOpnUPkzXonRwby8 

# Crosstalk glitch dataset
gdown 1h4CUR9rtw7wQRZZ4HGkShd3POhu00vzF

Installation

Install the required packages:

pip install -r requirements.txt

Model Training

We provide training scripts for crosstalk delay/glitch prediction. The repository includes both conventional GNN models and graph Transformer-based models for comparison.

Supported models include:

• GNN-based models: GCN, GAT, GIN, GraphSAGE, and DeepGCN
• Graph Transformer-based models: Graphomer, SGFormer, GraphGPS, and Si-GT

The training has two modes:

• segment: predict delay/glitch at the segment level
• sink: predict delay/glitch at the sink level

Set vic_only = True to use only nodes on victim net for training and evaluation.

Model Configuration

To change the GNN backbone, set the gnn_type parameter in the training script:

gnn_type = "gin"  # options: "gin", "gcn", "gat", "sage"

To enable positional encoding, set pe_type and provide the corresponding configuration in cfg. You can combining multiple positional encodings:

pe_type = ["EquivStableLapPE", "RWSE"]

cfg = {
    "posenc_RWSE": {
        "enable": True,
        "kernel": {
            "times_func": range(1, 17),
            "times": list(range(1, 17)),
        },
        "model": "Linear",
        "dim_pe": 16,
        "raw_norm_type": "BatchNorm",
    },
    "posenc_EquivStableLapPE": {
        "enable": True,
        "eigen": {
            "laplacian_norm": "none",
            "eigvec_norm": "L2",
            "max_freqs": 8,
        },
        "raw_norm_type": "none",
    },
}

Training Command

To train a delay/glitch prediction model, change model parameter in the script, then run:

python crosstalk_delay.py
python crosstalk_glitch.py

Model Testing

We provide two Jupyter notebooks for loading trained checkpoints, running inference, and evaluating model accuracy under different segment lengths:

• glitch_inference.ipynb
• delay_inference.ipynb

Citation

If you find this repository useful, please consider citing our paper:

@inproceedings{
hu2026sigt,
title={Si-{GT}: Fast Interconnect Signal Integrity Analysis for Integrated Circuit Design via Graph Transformers},
author={Yuting Hu and Tarek Mohamed and Chenhui Xu and Hua Xiang and Hussam Amrouch and Gi-Joon Nam and Jinjun Xiong},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=orO5727bSh}
}