MAGE AI Improves Chip Macro Placement with Human-Like Reasoning.

Chip design's physical layout, specifically macro placement, traditionally demands extensive manual effort. A new framework, MAGE (Macro Placement Agentic Engine), aims to automate and enhance this process using a multimodal, multi-agent approach. MAGE breaks down the complex placement task into a six-phase workflow, integrating established floorplanning rules with visual assessments and continuous refinement loops. Instead of learning from pre-labeled data, MAGE incorporates expert floorplanning knowledge through natural language directives and validation criteria. It also features a tournament-style evaluation system that compares multiple placement candidates and uses feedback from superior solutions to improve subsequent iterations. The framework introduces novel metrics like notch, whitespace, pocket, and alignment scores to quantify how human-like a placement is, capturing nuances often missed by standard performance metrics. Evaluations across various chip designs show MAGE achieving substantial improvements in critical timing metrics (WNS and TNS) over both commercial tools and human experts, while maintaining comparable wirelength and power. It also significantly boosts human-likeness scores, demonstrating its ability to generate layouts that align with expert design principles. The system's ability to transfer to new placement settings without specific retraining highlights its robustness.