Abstract
Accurate timing prediction is a cornerstone of modern integrated circuit (IC) design, yet high-fidelity timing information is traditionally unavailable until after the computationally intensive routing stage. This paper introduces PreRoutGNN, a novel two-stage Graph Neural Network (GNN) framework designed to predict post-routing timing metrics—such as slack and arrival time—directly from pre-routing placements. By integrating global circuit pre-training with a specialized Order Preserving Partition (OPP) scheme, PreRoutGNN achieves a significant leap in prediction accuracy, moving from a previous state-of-the-art $R^2$ of 0.59 to 0.93, while simultaneously reducing GPU memory requirements to manageable levels for industrial-scale designs.
Introduction
In the physical design flow of integrated circuits, "timing closure" determines whether a design can operate at its target frequency. The primary metric, slack, indicates whether signals arrive at their destination within the required clock cycle. Historically, designers face a catch-22: accurate timing analysis requires the final wire geometry provided by the routing stage, but waiting for routing to identify timing violations leads to costly, time-consuming design iterations.
Predicting these metrics pre-routing is the ideal solution, but it presents three major technical hurdles:
- Signal Decay: In deep logic paths, small errors in local delay prediction accumulate, leading to massive inaccuracies at the output.
- Global Context: Timing is not just a local phenomenon; it depends on the structural characteristics of the entire circuit graph.
- Computational Scale: Modern circuits contain millions of components, exceeding the memory capacity of standard hardware for deep learning training.
PreRoutGNN addresses these challenges through a hierarchical approach that combines global structural awareness with localized physical modeling.
Methodology
1. Global Circuit Pre-training
To capture the macro-scale features of a circuit, PreRoutGNN employs a self-supervised Stage 1. A Graph Auto-Encoder (GAE) is trained to reconstruct the circuit netlist, forcing the model to learn a compressed, "global view" embedding of the entire graph. These embeddings serve as a rich feature set for the subsequent prediction task, providing the context that local GNNs typically lack.
2. Residual Local Learning (RLL) and Attentional Modeling
For Stage 2, the model focuses on the physics of delay.
- Residual Local Learning: Instead of treating timing as a black-box regression, the model uses a residual scheme to explicitly calculate the incremental delay added by each net and cell along a path.
- Multi-head Joint Attention (MJA): The authors observed that traditional timing engines use Look-Up Tables (LUTs) to determine cell delay based on input slew and output load. PreRoutGNN mimics this by using an attention mechanism where slew and load act as "queries" and the LUT data acts as "keys" and "values," allowing the model to learn the underlying physics of the standard cell library.
3. Order Preserving Partition (OPP)
To handle massive designs without exhausting GPU memory, the framework introduces Order Preserving Partition. Unlike standard graph partitioning that may cut the graph randomly, OPP partitions the circuit based on its topological order—the actual sequence of signal flow. By adding "padding" layers to these sub-graphs, the model maintains the necessary context from preceding and succeeding logic levels, ensuring that the sequential nature of timing is preserved even when the graph is fragmented for training.
Experimental Results
The framework was evaluated using 21 real-world circuits from the OpenROAD benchmark, including complex designs like aes128 and xtea, with some designs exceeding 300,000 pins.
Accuracy and Efficiency
PreRoutGNN demonstrated a transformative improvement in accuracy. On the critical task of slack prediction, it achieved an $R^2$ score of 0.93, compared to the 0.59 achieved by the previous state-of-the-art, TimingGCN. This high correlation suggests that the model can reliably replace full routing for early-stage timing validation.
From a computational standpoint, the OPP scheme reduced peak GPU memory usage from over 48GB (which caused Out-of-Memory errors on standard hardware) to just 22GB. Furthermore, the model is highly performant; it can perform inference on a circuit with 300,000 pins in less than one second, enabling real-time feedback for designers.
Conclusion
PreRoutGNN represents a significant step forward in AI-driven Electronic Design Automation (EDA). By combining topological awareness with attentional mechanisms that mimic physical timing engines, it overcomes the traditional trade-off between speed and accuracy in pre-routing analysis. Future work will focus on integrating PreRoutGNN directly into timing-driven placement engines to automate the optimization of chip layouts, further shortening the semiconductor design cycle.
Summary updated