Abstract:The Rectilinear Steiner Minimum Tree (RSMT) problem is a fundamental problem in VLSI placement and routing and is known to be NP-hard. Traditional RSMT algorithms spend a significant amount of time on finding Steiner points to reduce the total wire length or use heuristics to approximate producing sub-optimal results. We show that Graph Neural Networks (GNNs) can be used to predict optimal Steiner points in RSMTs with high accuracy and can be parallelized on GPUs. In this paper, we propose GAT-Steiner, a graph attention network model that correctly predicts 99.846% of the nets in the ISPD19 benchmark with an average increase in wire length of only 0.480% on suboptimal wire length nets. On randomly generated benchmarks, GAT-Steiner correctly predicts 99.942% with an average increase in wire length of only 0.420% on suboptimal wire length nets.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: in VLSI (Very - Large - Scale Integration) placement and routing, how to solve the Rectilinear Steiner Minimum Tree (RSMT) problem efficiently and accurately. The RSMT problem is NP - hard. Traditional algorithms take a great deal of time when looking for Steiner points to reduce the total wire length, or use heuristic methods to obtain sub - optimal results. Specifically, the paper proposes a model based on Graph Neural Networks (GNNs) - GAT - Steiner, which is used to predict Steiner points in RSMT. Through this method, multiple instances can be processed in parallel on the GPU, thus significantly improving the computational efficiency, and being able to predict Steiner points with high precision, making the increase in the average wire length very small. ### Description of the core problems in the paper: 1. **Problem background**: - The RSMT problem is crucial in VLSI placement and routing, but it is NP - hard. - Traditional methods such as FLUTE and SALT can quickly find approximate solutions, but sacrifice the optimality of wire length. - Existing exact solution methods such as GeoSteiner are accurate but take too much time and are difficult to be applied to large - scale design optimization. 2. **Solution**: - Propose the GAT - Steiner model, which uses the Graph Attention Network (GAT) to directly predict Steiner points. - This model can perform parallel computing on the GPU, significantly reducing the running time. - Experiments show that GAT - Steiner correctly predicts 99.846% of the networks in the ISPD19 benchmark test, with an average wire length increase of only 0.480%, and performs even better in randomly generated benchmark tests. 3. **Advantages**: - High precision: It performs well on multiple test sets, especially in randomly generated large - scale networks. - Parallelization: It can efficiently process multiple instances in parallel on the GPU, significantly accelerating the computation. - Reduction of outliers: Compared with heuristic methods, the number of wire length outliers generated by GAT - Steiner is reduced by 1 - 2 orders of magnitude. Through these improvements, GAT - Steiner provides an efficient and accurate new method for solving the RSMT problem, especially suitable for design processes that require rapid iteration and optimization.

GAT-Steiner: Rectilinear Steiner Minimal Tree Prediction Using GNNs

NN-Steiner: A Mixed Neural-algorithmic Approach for the Rectilinear Steiner Minimum Tree Problem

RLGAT: Retweet Prediction in Social Networks Using Representation Learning and GATs

GATR: A Road Network Traffic Violation Prediction Method Based on Graph Attention Network

NeuroSteiner: A Graph Transformer for Wirelength Estimation

GPU-Accelerated Rectilinear Steiner Tree Generation.

NTGAT: A Graph Attention Network Accelerator with Runtime Node Tailoring

Pre-Training Identification of Graph Winning Tickets in Adaptive Spatial-Temporal Graph Neural Networks

How Attentive are Graph Attention Networks?

Spiking GATs: Learning Graph Attentions via Spiking Neural Network

Nearly Optimal Steiner Trees using Graph Neural Network Assisted Monte Carlo Tree Search

Predicting future production system bottlenecks with a graph neural network approach

Neural Trees for Learning on Graphs

JITGNN: A deep graph neural network framework for Just-In-Time bug prediction

SH-GAT: Software-hardware co-design for accelerating graph attention networks on FPGA

GTuner

Breaking the Bottleneck on Graphs with Structured State Spaces

Graph Neural Network Training with Data Tiering

Adaptive Depth Graph Attention Networks

RATs-NAS: Redirection of Adjacent Trails on GCN for Neural Architecture Search

Optimization and Interpretability of Graph Attention Networks for Small Sparse Graph Structures in Automotive Applications