Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ChatEDA: A Large Language Model Powered Autonomous Agent for EDA

Zhuolun He, Haoyuan Wu, Xinyun Zhang, Xufeng Yao, Su Zheng, Haisheng Zheng, Bei Yu

TL;DR

The paper introduces ChatEDA, an autonomous agent that leverages a domain-focused, fine-tuned LLM (AutoMage) to orchestrate RTL-to-GDSII EDA workflows through external tool APIs. It details principled methods—LoRA/QLoRA-based efficient fine-tuning, in-context learning, and instruction tuning with explanations and chain-of-thought—to build robust, expert EDA controllers, enhanced in AutoMage2 with enriched data and reasoning prompts. Experimental evaluation via ChatEDA-Bench shows AutoMage2 achieving superior task decomposition and script generation, notably surpassing GPT-4 in Grade-A outcomes (82% vs 62%), indicating strong practical impact for automating complex EDA toolchains. The work highlights both the potential and limitations of LLM-based EDA automation, emphasizing the need for broader tool compatibility, scalable evaluation, and improved generalization for real-world deployment.

Abstract

The integration of a complex set of Electronic Design Automation (EDA) tools to enhance interoperability is a critical concern for circuit designers. Recent advancements in large language models (LLMs) have showcased their exceptional capabilities in natural language processing and comprehension, offering a novel approach to interfacing with EDA tools. This research paper introduces ChatEDA, an autonomous agent for EDA empowered by an LLM, AutoMage, complemented by EDA tools serving as executors. ChatEDA streamlines the design flow from the Register-Transfer Level (RTL) to the Graphic Data System Version II (GDSII) by effectively managing task decomposition, script generation, and task execution. Through comprehensive experimental evaluations, ChatEDA has demonstrated its proficiency in handling diverse requirements, and our fine-tuned AutoMage model has exhibited superior performance compared to GPT-4 and other similar LLMs.

ChatEDA: A Large Language Model Powered Autonomous Agent for EDA

TL;DR

The paper introduces ChatEDA, an autonomous agent that leverages a domain-focused, fine-tuned LLM (AutoMage) to orchestrate RTL-to-GDSII EDA workflows through external tool APIs. It details principled methods—LoRA/QLoRA-based efficient fine-tuning, in-context learning, and instruction tuning with explanations and chain-of-thought—to build robust, expert EDA controllers, enhanced in AutoMage2 with enriched data and reasoning prompts. Experimental evaluation via ChatEDA-Bench shows AutoMage2 achieving superior task decomposition and script generation, notably surpassing GPT-4 in Grade-A outcomes (82% vs 62%), indicating strong practical impact for automating complex EDA toolchains. The work highlights both the potential and limitations of LLM-based EDA automation, emphasizing the need for broader tool compatibility, scalable evaluation, and improved generalization for real-world deployment.

Abstract

The integration of a complex set of Electronic Design Automation (EDA) tools to enhance interoperability is a critical concern for circuit designers. Recent advancements in large language models (LLMs) have showcased their exceptional capabilities in natural language processing and comprehension, offering a novel approach to interfacing with EDA tools. This research paper introduces ChatEDA, an autonomous agent for EDA empowered by an LLM, AutoMage, complemented by EDA tools serving as executors. ChatEDA streamlines the design flow from the Register-Transfer Level (RTL) to the Graphic Data System Version II (GDSII) by effectively managing task decomposition, script generation, and task execution. Through comprehensive experimental evaluations, ChatEDA has demonstrated its proficiency in handling diverse requirements, and our fine-tuned AutoMage model has exhibited superior performance compared to GPT-4 and other similar LLMs.
Paper Structure (26 sections, 6 equations, 5 figures)

This paper contains 26 sections, 6 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Generative Pre-Trained Language Model
  4. Low-Rank Adaptation of LLMs
  5. In-Context Learning
  6. Block-wise k-bit Quantization
  7. ChatEDA: LLM-powered Framework for EDA
  8. Task Decomposition
  9. Script Generation
  10. Task Execution
  11. AutoMage: LLM-based Controller of ChatEDA
  12. Base Model Selection
  13. Instruction Tuning for AutoMage.
  14. Efficient Fine-tuning of Quantized LLMs
  15. Auto-Regressive Decoding
  16. ...and 11 more sections

Figures (5)

  • Figure 1: Overview of AutoMage powered ChatEDA. With AutoMage as the controller and EDA tools as the executors, the workflow consists of three stages: 1) Task Decomposition; 2) Script Generation; 3) Task Execution.
  • Figure 2: Language functions as a conduit enabling ChatEDA to integrate EDA tools for resolving complex EDA tasks. Within the framework, ChatEDA acts as the controller that harmonizes and orchestrates the collaboration among various tools. ChatEDA first formulates a task list derived from user requirements, subsequently generating scripts corresponding to these decomposed tasks.
  • Figure 3: Overview of Instruction Tuning. During the instruction tuning process, we use the self instruction paradigm to construct our instruction pool via GPT models. Then we apply the QLoRA technique for efficient instruction fine-tuning.
  • Figure 4: Examples of generated EDA tools instructions. Moreover, we also provide more examples in the repo https://github.com/wuhy68/ChatEDAv1 for a better understanding of the generated dataset.
  • Figure 5: Evaluation results for AutoMage he2023chateda and AutoMage2 compared to other LLMs. AutoMage models outperform other notable LLMs by a significant margin in task planning and script generation and AutoMage2 performs the best.