[Paper Review] Net2: A Graph Attention Network Method Customized for Pre-Placement Net Length Estimation

Zhiyao Xie, Rongjian Liang+4 more|arXiv (Cornell University)|Nov 27, 2020

VLSI and FPGA Design Techniques24 references6 citations

TL;DR

This paper proposes Net², a customized graph attention network (GAT) method for pre-placement net length estimation in VLSI design. By leveraging global circuit topology through circuit partitioning and attention mechanisms, Net² achieves 15% higher accuracy than prior methods in identifying long nets and critical paths, with a fast version over 1000× faster than placement.

ABSTRACT

Net length is a key proxy metric for optimizing timing and power across various stages of a standard digital design flow. However, the bulk of net length information is not available until cell placement, and hence it is a significant challenge to explicitly consider net length optimization in design stages prior to placement, such as logic synthesis. This work addresses this challenge by proposing a graph attention network method with customization, called Net2, to estimate individual net length before cell placement. Its accuracy-oriented version Net2a achieves about 15% better accuracy than several previous works in identifying both long nets and long critical paths. Its fast version Net2f is more than 1000 times faster than placement while still outperforms previous works and other neural network techniques in terms of various accuracy metrics.

Motivation & Objective

• Address the challenge of estimating individual net lengths before cell placement, a critical need for timing and power optimization in early design stages.
• Overcome the limitation of prior methods that rely only on local topology and lack global context in net length prediction.
• Enable accurate and efficient net length estimation to support layout-aware logic synthesis and early-stage optimization.
• Develop a deep learning approach that captures long-range dependencies in netlist graphs using graph attention networks with customized features and partitioning.
• Balance accuracy and speed by introducing two variants: Net²a (accuracy-focused) and Net²f (fast inference).

Proposed method

• Construct a netlist graph where nodes represent cells and edges represent nets, with edge features encoding topological and structural information.
• Apply a graph attention network (GAT) to aggregate node representations using learnable attention coefficients, enabling dynamic weighting of neighboring nodes.
• Introduce a circuit partitioning strategy to extract global information: the design is divided into clusters, and inter-cluster edges are used to model long-range dependencies.
• Use edge convolution layers to process edge features (F0–F3) that encode net degree, cell proximity, and topological distance, enhancing feature representation.
• Train the model end-to-end using a regression loss to predict actual net lengths from pre-placement netlists.
• Deploy two variants: Net²f for speed (no partitioning, single GAT layer), and Net²a for accuracy (with partitioning and deeper attention layers).

Experimental results

Research questions

• RQ1 Can a graph neural network effectively estimate individual net lengths before placement, capturing both local and global topology?
• RQ2 How does incorporating global circuit structure through partitioning improve net length estimation accuracy compared to local-only models?
• RQ3 To what extent does attention-based message passing outperform traditional GNNs or fully connected networks in this context?
• RQ4 What is the trade-off between inference speed and accuracy in pre-placement net length prediction, and can a model be both fast and accurate?
• RQ5 How do different edge features (F0–F3) contribute to the predictive performance of the model?

Key findings

• Net²a achieves 15% higher accuracy than previous methods in identifying the 10% longest nets, with an average ROC AUC of 92.2% across benchmarks.
• Net²f is over 1000× faster than standard placement while maintaining superior accuracy, making it suitable for real-time optimization.
• The inclusion of F1 edge features (representing net connectivity patterns) contributes most significantly to accuracy improvement.
• Using coarser cell partitions (less P) reduces inference time by ~1.5× and maintains 91.6% of Net²a’s accuracy, indicating robustness to partition granularity.
• The model without GNN (Edge ANN) performs worst, confirming that message passing via GAT is essential for capturing structural context.
• Net²a is more than 10× faster than placement, and training the model takes approximately 30 minutes on standard benchmarks.