跳转至

文章背景与核心概要

逻辑综合是将 RTL 设计转换为门级网表的核心步骤，但其功耗、性能和面积（PPA）结果对优化指令极其敏感，这使得综合调优成为一个高维且高成本的难题。传统的调优方法要么依赖缺乏可解释性的黑盒自动化搜索，要么采用无法适应不断变化的电路状态的静态大语言模型（LLM）脚本。

为了克服这些局限性，作者推出了 SynAct——一个自适应的闭环 LLM 推理-行动（Reasoning-Acting）智能体。SynAct 能够动态诊断实时的综合报告，对当前电路状态进行推理，检索相关的工具知识，并利用历史优化经验来下发针对性的指令。通过重点改善时序（特别是最差负裕量 WNS）并在面积与功耗之间保持平衡，SynAct 在 14 个商业工具设计中，将平均 WNS 显著降低至基础引导综合（baseline bootstrap synthesis）的 27%。

SynAct: A Reasoning-Acting Large Language Model Agent for Adaptive Synthesis Optimization

arXiv: 2608.12751 [cs.AR]
Submitted: August 13, 2026
Subjects: Hardware Architecture (cs.AR); Artificial Intelligence (cs.AI)
Authors: Fangzhou Liu, Peiyi Han, Jiawei Liu, Yuan Pu, Zhuolun He, Rongliang Fu, Tsung-Yi Ho, Bei Yu

📌 摘要与总结

Logic synthesis transforms RTL designs into gate-level netlists, but Power, Performance, and Area (PPA) outcomes are exceptionally sensitive to optimization commands, making synthesis tuning a high-dimensional and expensive challenge. Traditional approaches rely either on black-box automated search methods with limited interpretability, or static Large Language Model (LLM) scripts that cannot adapt to evolving circuit states.

逻辑综合将 RTL 设计转换为门级网表，但功耗、性能和面积（PPA）结果对优化指令极其敏感，这使得综合调优成为一个高维且代价高昂的挑战。传统方法要么依赖于可解释性有限的黑盒自动化搜索方法，要么依赖于无法适应演变中的电路状态的静态大语言模型（LLM）脚本。

To overcome these limitations, the authors introduce SynAct, an adaptive closed-loop LLM reasoning-acting agent. SynAct dynamically diagnoses live synthesis reports, reasons over the current circuit state, retrieves relevant tool knowledge, and leverages historical optimization experience to issue targeted commands. By focusing heavily on improving timing—specifically worst negative slack (WNS)—while maintaining balanced area and power trade-offs, SynAct achieved a reduction of average WNS down to 27% of baseline bootstrap synthesis across 14 commercial tool designs.

为了克服这些局限性，作者推出了 SynAct，这是一个自适应的闭环 LLM 推理-行动智能体。SynAct 动态诊断实时的综合报告，推导当前的电路状态，检索相关的工具知识，并利用历史优化经验来下发针对性的指令。通过专注于改善时序（特别是最差负裕量 WNS）同时保持面积和功耗的平衡权衡，SynAct 在 14 个商业工具设计中将平均 WNS 降低至基础引导综合的 27%。

👥 作者团队

  • Fangzhou Liu
  • Peiyi Han
  • Jiawei Liu
  • Yuan Pu
  • Zhuolun He
  • Rongliang Fu
  • Tsung-Yi Ho
  • Bei Yu

👥 Authors

  • Fangzhou Liu
  • Peiyi Han
  • Jiawei Liu
  • Yuan Pu
  • Zhuolun He
  • Rongliang Fu
  • Tsung-Yi Ho
  • Bei Yu

📄 论文摘要

Logic synthesis transforms RTL designs into gate-level netlists, where PPA results are highly sensitive to the choice of optimization commands, making synthesis tuning both high-dimensional and expensive. Previous approaches fall into two categories: automated methods, which perform black-box search over fixed action spaces with limited decision-level interpretability, and LLM-based methods, which typically generate static scripts upfront and cannot adapt to evolving circuit states. We present SynAct, an adaptive closed-loop LLM reasoning–acting agent that iteratively diagnoses live synthesis reports and reasons over the current circuit state, retrieved tool knowledge, and historical optimization experience to issue targeted commands. SynAct focuses on improving timing, particularly worst negative slack (WNS), while maintaining balanced area and power trade-offs. Experiments on a commercial synthesis tool across 14 designs show that SynAct reduces average WNS to 27% of that from bootstrap synthesis.

逻辑综合将 RTL 设计转换为门级网表，其中 PPA 结果对优化命令的选择高度敏感，这使得综合调优既具有高维度又成本高昂。先前的方法分为两类：一是自动化方法，在固定的动作空间上进行黑盒搜索，决策层面的可解释性有限；二是基于 LLM 的方法，通常在前期生成静态脚本，无法适应演变中的电路状态。我们提出了 SynAct，一个自适应的闭环 LLM 推理-行动智能体，它能够迭代地诊断实时综合报告，并结合当前电路状态、检索到的工具知识以及历史优化经验进行推理，从而下发有针对性的命令。SynAct 专注于改善时序，特别是最差负裕量（WNS），同时在面积和功耗之间保持平衡。在商业综合工具上对 14 个设计的实验表明，SynAct 将平均 WNS 降低到了引导综合结果的 27%。

🔗 资源与链接