GR-Evolve: Design-Adaptive Global Routing via LLM-Driven Algorithm Evolution

Overview

GR-Evolve is a research framework that uses Large Language Models (LLMs) to automatically evolve global routing algorithms tailored to specific circuit designs. Instead of hand-crafting heuristics or training black-box ML models, GR-Evolve treats algorithm design itself as the search space: an LLM proposes algorithmic variants, candidates are evaluated on real routing benchmarks, and the population evolves toward solutions that adapt to the structural characteristics of each design.

This repository contains the official implementation of the paper:

For a full description of the methodology, motivation, and experimental results, please refer to the paper.

Repository Structure

GR-Evolve/
├── AGENTS                # Contains AGENTS.md file for each benchmark and tech node.
│   └── *.md  
│   
├── Dockerfile            # Dockerfile to build docker image to run multiple benchmarks in parallel
├── GR_SUMMARY.md         # Global routing paper summary. 
├── OpenROAD-flow-scripts # ORFS-Suite used to run GR+DR flow
├── OpenROAD_New_GRT      # OpenROAD fork where LLM will make source code changes
├── README.md             # This file
└── SHELL_SCRIPTS         # Directory containg shell scripts for LLM to use
    └── *.md  

Installation

Prerequisites

• An API key for your preferred LLM provider (e.g., Anthropic, OpenAI)

Setup

We use docker to create our router-design evolution. The Docker workflow runs evolution in parallel across multiple benchmarks and technology nodes, with one container per (router × design × PDK) combination.

To use docker to run multiple design-router pair evolutions in parallel, please run the runevolve.sh script using:

bash SHELL_SCRIPTS/runevolve.sh 

This will create the 45 router-design pairs used in our experiments.

What the setup script does

The runevolve.sh script iterates over every container in the list and performs four steps automatically:

Step 1 — Container setup ([DOCKER]) Creates the Docker container from the built image, copies the SHELL_SCRIPTS/ directory into it, provides the GR_SUMMARY.md reference document, and seeds the METRICS_TABLE.md file used to record Quality-of-Results (QoR) metrics. It also checks out the correct source branch for the router assigned to that container (FastRoute, CUGR, or SPRoute).

Step 2 — Codex smoke test ([TEST CODEX]) Runs a lightweight sanity check to confirm that the Codex agent inside the container has the correct write permissions. It instructs Codex to create a folder with a four-column Markdown file. If this step succeeds, the agent is ready to make source code changes.

Step 3 — AGENTS file setup ([SETUP EVOLUTION]) Copies the correct AGENTS.md file from the AGENTS/ directory into /root/AGENTS.md inside the container. The file is selected automatically based on the design and PDK encoded in the container name (e.g., fr___aes_sky130 receives AGENTS_AES_sky130.md). This file provides the LLM agent with design-specific context and instructions.

Step 4 — Start evolution ([START EVOLUTION]) Launches GeneticRunCodex.sh in the background via nohup. Evolution runs are logged to /root/evolution_start_time.log inside each container. From this point the containers run autonomously.