Yunqi He

Yunqi He

Ph.D. Candidate in Computer Engineering — Physical Design Methodology & ML for EDA

Project Duration

2025 - Present

Collaboration

NuLogiCS Research Group; Cadence Design Systems

Keywords

Graph Neural Networks, Timing-Driven Placement, Congestion Prediction, Floorplanning

ML for Physical Design

Published:

Project Overview

This project applies graph neural networks to core physical-design tasks, learning cell-, net-, and layout-level representations directly from the netlist rather than relying on hand-crafted heuristics. The goal is to move PPA where it matters — timing, routability, and convergence — by giving classical engines better starting points and better predictions.

Timing-Driven Placement Initialization

Graph Neural Network Based Initialization for Timing Driven Placement (TODAES 2026) frames placement initialization as node-level prediction on the netlist graph.

  • GNNs learn cell- and net-level embeddings that produce timing-aware initial placements.
  • The learned seed warm-starts the analytical placer, cutting placement iterations and improving post-placement timing over conventional initialization.
  • Improves downstream timing and convergence without changing the underlying placer.

Routing-Congestion Prediction (Industry)

As a returning R&D intern at Cadence Design Systems, this work builds GNN-based routing congestion prediction for the Palladium (Z4) emulator — forecasting congestion from netlist and placement features to guide routing and improve routability on industrial designs, extending prior GPU placement-acceleration work into the routing stage.

Data-Driven Floorplanning

The GNN line is being extended to floorplanning: a data-driven, agile ML-based floorplanning tool that learns from the FloorSet dataset of real-world SoC layouts to guide block placement. Paired with the TwinTree legalizer, it serves as the floorplanning engine within the dissertation’s end-to-end agentic flow.

  1. Graph Neural Network Based Initialization for Timing Driven Placement — TODAES 2026

Future Directions

  • Unified GNN representations shared across placement, floorplanning, and routing stages.
  • Congestion- and timing-aware learning integrated directly into GPU-accelerated engines.
  • ML engines exposed as callable primitives for the agentic design-automation system.