Learning to Place Macros and Synthesize Power Grids Through Multi Agent Coordination

Abstract

Modern chip design faces unprecedented challenges in optimizing macro placement and power grid synthesis simultaneously. Traditional Electronic Design Automation (EDA) approaches rely on sequential optimization strategies that fail to capture the complex interdependencies between macro positioning and power distribution networks. This paper presents a novel framework leveraging multi-agent reinforcement learning for coordinated macro placement and power grid synthesis. Our approach employs multiple specialized agents that collaboratively optimize placement objectives including wirelength minimization, congestion reduction, and power integrity. Through a hierarchical coordination mechanism, agents negotiate placement decisions while maintaining awareness of power delivery constraints. Experimental results on industrial benchmarks demonstrate that our multi-agent coordination framework achieves 12.3% improvement in wirelength, 15.7% reduction in congestion hotspots, and 18.2% enhancement in IR drop metrics compared to conventional single-agent and sequential optimization methods. The framework exhibits strong scalability properties, handling designs with over 1000 macros while maintaining solution quality. This work demonstrates that multi-agent coordination provides a promising paradigm for addressing the increasing complexity of modern chip design challenges.