Research Papers 论文研究 5h ago Updated 1h ago 更新于 1小时前 46

MacroAgent: Regularity-Aware Macro Legalization with LLM-Agent-Designed Contour Algorithms MacroAgent:基于LLM智能体设计轮廓算法的规律感知宏布局合法化

MacroAgent is a novel four-stage macro legalization framework for VLSI designs that leverages LLMs to discover heuristic regularity-aware contour algorithms The approach achieves 2-8 fold improvement in layout regularity and 3-5% reduction in routed wirelength compared to state-of-the-art methods on TILOS and Chipyard benchmarks End-to-end evaluation via Cadence Innovus confirms tangible PPA gains: 2.9% lower routed wirelength and 68.3% TNS improvement over DREAMPlace baseline The framework addr 提出MacroAgent框架,利用LLM自动发现启发式轮廓算法解决VLSI宏单元合法化问题 四阶段流水线(聚类→轮廓生成→模板匹配→簇间优化)显著提升布局规律性2-8倍 在TILOS和Chipyard基准上布线长度减少3%-5%,Cadence Innovus端到端评估TNS改善68.3% 相比DREAMPlace基线实现2.9%布线长度降低,验证了规律感知策略的PPA收益

62
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • MacroAgent is a novel four-stage macro legalization framework for VLSI designs that leverages LLMs to discover heuristic regularity-aware contour algorithms
  • The approach achieves 2-8 fold improvement in layout regularity and 3-5% reduction in routed wirelength compared to state-of-the-art methods on TILOS and Chipyard benchmarks
  • End-to-end evaluation via Cadence Innovus confirms tangible PPA gains: 2.9% lower routed wirelength and 68.3% TNS improvement over DREAMPlace baseline
  • The framework addresses key limitations of existing approaches—lack of robustness, high computational cost, and neglect of inter-macro regularity
  • LLM-agent-designed contour algorithms enable robust, effective heuristic solutions with acceptable runtime performance

Why It Matters

This work represents a significant intersection of large language models and electronic design automation (EDA), demonstrating that LLMs can autonomously discover effective algorithmic heuristics for a critical physical design problem. For AI practitioners and EDA researchers, it showcases a novel paradigm where LLMs act as algorithm designers rather than mere code assistants, with direct impact on chip performance, power, and area (PPA) metrics that matter to the semiconductor industry.

Technical Details

  • Four-stage pipeline: The framework consists of clustering, contour generation, template matching, and inter-cluster refinement, addressing macro legalization as a structured optimization problem
  • LLM-driven algorithm discovery: Large Language Models are employed to autonomously generate multiple effective heuristic regularity-aware contour algorithms, rather than relying on hand-crafted heuristics
  • Regularity-aware design: Unlike prior work, MacroAgent explicitly considers regularity between macros, which is critical for manufacturability and downstream routing quality
  • Benchmark validation: Evaluated on TILOS and Chipyard benchmarks, showing 2-8x layout regularity improvement, 3-5% routed wirelength reduction, and comparable congestion after global routing
  • Industry-grade evaluation: End-to-end testing through Cadence Innovus place-and-route flow confirmed 2.9% lower routed wirelength and 68.3% total negative slack (TNS) improvement over DREAMPlace, with 1.8% wirelength reduction when integrated into the Innovus macro placement flow

Industry Insight

  • The successful application of LLMs as autonomous algorithm designers in EDA signals a broader trend: domain-specific heuristic discovery through generative AI will become a competitive differentiator in chip design toolchains
  • The tangible PPA improvements validated through industry-standard Cadence Innovus flows demonstrate that AI-driven EDA solutions can move beyond academic benchmarks to real-world deployment, accelerating adoption timelines
  • Foundries and fabless companies should monitor this approach as a potential replacement or augmentation to existing legalization flows, particularly for large-scale designs where macro regularity directly impacts yield and performance

TL;DR

  • 提出MacroAgent框架,利用LLM自动发现启发式轮廓算法解决VLSI宏单元合法化问题
  • 四阶段流水线(聚类→轮廓生成→模板匹配→簇间优化)显著提升布局规律性2-8倍
  • 在TILOS和Chipyard基准上布线长度减少3%-5%,Cadence Innovus端到端评估TNS改善68.3%
  • 相比DREAMPlace基线实现2.9%布线长度降低,验证了规律感知策略的PPA收益

为什么值得看

该工作展示了LLM在传统EDA领域的创新应用,为宏单元布局优化提供了自动化算法发现的新范式。规律感知策略不仅提升布局质量,还通过Cadence工业流程验证了实际PPA收益,对芯片设计自动化具有参考价值。

技术解析

  • 框架架构:四阶段方法包括聚类、轮廓生成、模板匹配和簇间优化,形成完整的宏单元合法化流程
  • LLM应用:利用大语言模型自动发现多种有效的启发式规律感知轮廓算法,替代人工设计
  • 实验基准:在TILOS和Chipyard标准测试集上验证,对比SOTA方法
  • 性能指标:布局规律性提升2-8倍,布线长度减少3%-5%,全局路由后拥塞相当
  • 工业验证:通过Cadence Innovus place-and-route端到端评估,相比DREAMPlace基线布线长度降低2.9%,TNS改善68.3%

行业启示

  • LLM-Agent范式可拓展至传统EDA领域,为硬件设计自动化提供新的算法发现路径
  • 规律性感知成为提升芯片物理设计质量的关键方向,需在设计流程中显式建模
  • 工业级验证(Cadence Innovus集成)表明AI方法已具备落地潜力,建议关注LLM与EDA工具的深度融合趋势

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Agent Agent Research 科学研究 Chip 芯片 Programming 编程