Divergent Thoughts toward One Goal: LLM-based Multi-Agent Collaboration System for Electronic Design Automation

TL;DR

The paper tackles the challenge of automating complex EDA flows across diverse tool interfaces by introducing EDAid, a multi-agent system that leverages ChipLlama expert LLMs to maintain reliable, long-chain tool-calling. It combines a single-agent ChipLlama with hybrid instruction tuning and few-shot CoT prompts, and a multi-agent framework where divergent-thoughts agents propose multiple planning pathways that are reconciled by a decision-making agent. Empirical results on ChatEDA-bench and iEDA-bench show state-of-the-art performance, with ChipLlama-70B achieving near-perfect accuracies (reported as 100% on both benchmarks) and EDAid outperforming single-agent baselines across platforms. The work demonstrates a scalable, cross-platform approach to automate EDA flows, reducing intermediate errors and enabling practical deployment, though it acknowledges higher inference latency as a trade-off and points to optimization opportunities for real-world use.

Abstract

Recently, with the development of tool-calling capabilities in large language models (LLMs), these models have demonstrated significant potential for automating electronic design automation (EDA) flows by interacting with EDA tool APIs via EDA scripts. However, considering the limited understanding of EDA tools, LLMs face challenges in practical scenarios where diverse interfaces of EDA tools exist across different platforms. Additionally, EDA flow automation often involves intricate, long-chain tool-calling processes, increasing the likelihood of errors in intermediate steps. Any errors will lead to the instability and failure of EDA flow automation. To address these challenges, we introduce EDAid, a multi-agent collaboration system where multiple agents harboring divergent thoughts converge towards a common goal, ensuring reliable and successful EDA flow automation. Specifically, each agent is controlled by ChipLlama models, which are expert LLMs fine-tuned for EDA flow automation. Our experiments demonstrate the state-of-the-art (SOTA) performance of our ChipLlama models and validate the effectiveness of our EDAid in the automation of complex EDA flows, showcasing superior performance compared to single-agent systems.

Figures (8)

• Figure 1: Overview of the ChipLlama-powered agent for task planning and EDA script generation.
• Figure 2: Overview of hybrid instruction tuning.
• Figure 3: Few-shot CoT prompt template utilized in ChipLlama-powered agent.
• Figure 4: Overview of EDAid, the multi-agent collaboration system. Given an EDA task, multiple agents (including divergent-thoughts agents (role $R_{0}$) and a decision-making agent (role $R_{1}$)) collaborate to generate the EDA script. Finally, the generated EDA script will automate the EDA flow interfacing the EDA tools via APIs.
• Figure 5: Examples of evaluation benchmarks.