Painting on Placement: Forecasting Routing Congestion using Conditional Generative Adversarial Nets

Cunxi Yu, Zhiru Zhang

Abstract

Physical design process commonly consumes hours to days for large designs, and routing is known as the most critical step. Demands for accurate routing quality prediction raise to a new level to accelerate hardware innovation with advanced technology nodes. This work presents an approach that forecasts the density of all routing channels over the entire floorplan, with features collected up to placement, using conditional GANs. Specifically, forecasting the routing congestion is constructed as an image translation (colorization) problem. The proposed approach is applied to a) placement exploration for minimum congestion, b) constrained placement exploration and c) forecasting congestion in real-time during incremental placement, using eight designs targeting a fixed FPGA architecture.

Introduction

In the field of Electronic Design Automation (EDA), the "Place and Route" (PnR) process is a notorious bottleneck. While placement determines the physical locations of logic blocks, routing—the subsequent step of connecting them—is the most time-consuming and unpredictable phase. A design that appears optimal during placement may fail during routing due to excessive congestion, forcing engineers into a costly "design closure" loop that can last hours or even days.

This paper addresses the need for early, accurate congestion forecasting. By providing a fast prediction immediately after placement, designers can iterate rapidly and avoid wasting resources on unroutable placements. The authors propose treating this challenge not as a traditional regression problem, but as an image-to-image translation task.

Methodology: Congestion Prediction as Image Translation

The core innovation lies in framing routing congestion forecasting as a pix2pix style translation problem. Because a post-placement layout and its corresponding routing heat map share the same spatial geometry, the task is analogous to colorizing a grayscale image or translating a map to a satellite view.

The data generation pipeline, illustrating how features and ground truth routing maps are derived from the mapped netlist.

The researchers utilize a Conditional Generative Adversarial Network (cGAN). Unlike standard GANs that generate images from random noise, a cGAN is conditioned on specific input features. The architecture consists of a Generator (a Fully Convolutional Network with skip connections to preserve fine-grained spatial details) and a Discriminator trained to distinguish between authentic routing heat maps and those synthesized by the generator.

Visualizing the image translation task: from placement maps (a, b) to ground truth routing congestion heat maps (d).

Data and Feature Engineering

The study utilizes the VTR 8.0 (Verilog-to-Routing) open-source CAD flow. Training data was generated by sweeping placement parameters across eight FPGA benchmarks, resulting in a dataset of 1,500 input-output image pairs.

A critical finding was that spatial placement alone is insufficient for accurate prediction. The model utilizes a 4-channel input:

  • RGB Placement Image: Color-coded locations of logic blocks (CLBs), I/O pads, and specialized blocks (RAM/Multipliers).
  • Connectivity Image: A grayscale channel representing the netlist's edges drawn between block locations.

Including connectivity information significantly improved the model's ability to identify potential "traffic jams" that simple placement maps might miss.

Key Insights and Results

  • Real-Time Performance: The cGAN generates predictions in approximately 0.09 seconds, a massive speedup compared to the hours required for full routing.
  • High Fidelity: The model achieved per-pixel accuracy between 61% and 85%. More importantly, it proved highly effective at ranking placement candidates, identifying the least congested solutions with up to 90% accuracy.
  • Architectural Awareness: Using RGB images instead of grayscale improved accuracy by 3-5%, as the color helps the network distinguish between different physical constraints of the FPGA architecture.
  • Interactive Design: The speed of the model allows for real-time visualization of congestion during incremental placement, enabling a "live" view of how moving a block affects the overall routability of the chip.

Conclusion

By treating routing congestion as a visual translation problem, this work demonstrates that cGANs can provide near-instantaneous, high-fidelity feedback for FPGA physical design. While currently optimized for a fixed architecture, the approach opens the door for real-time design exploration and significantly faster hardware development cycles. Future work will focus on improving generalization across diverse circuit structures and adapting the model to varying hardware architectures.

Summary updated

Abstract

Physical design process commonly consumes hours to days for large designs, and routing is known as the most critical step. Demands for accurate routing quality prediction raise to a new level to accelerate hardware innovation with advanced technology nodes. This work presents an approach that forecasts the density of all routing channels over the entire floorplan, with features collected up to placement, using conditional GANs. Specifically, forecasting the routing congestion is constructed as an image translation (colorization) problem. The proposed approach is applied to a) placement exploration for minimum congestion, b) constrained placement exploration and c) forecasting congestion in real-time during incremental placement, using eight designs targeting a fixed FPGA architecture.
BibTeX
@inproceedings{yu2019painting,
  author = {Cunxi Yu and Zhiru Zhang},
  title = {{Painting on Placement: Forecasting Routing Congestion using Conditional Generative Adversarial Nets}},
  booktitle = {DAC},
  year = {2019},
  url = {https://openreview.net/forum?id=waL13OyL9JA},
}
Figure