Technical Summary: MAGE: Human-Like Macro Placement via Agentic Multimodal Reasoning

1. Problem Statement

Macro placement remains a critical bottleneck in VLSI physical design, often requiring substantial manual refinement by expert engineers to achieve high-quality results. While existing automated methods (classical optimization, analytical placers, and learning-based approaches) can produce legal placements with competitive Power, Performance, and Area (PPA) metrics, they frequently fail to reproduce the higher-level spatial regularity, functional coherence, and structural organization characteristic of expert-designed floorplans. Consequently, automated placements often lack the "human-like" qualities—such as coherent macro grouping, boundary alignment, and usable whitespace—that facilitate downstream optimization, routing, and timing closure.

2. Methodology: The MAGE Framework

The authors propose MAGE (Macro Placement Agentic Engine), a multimodal multi-agent framework designed to generate human-like macro placements without relying on labeled placement data or per-design training. Instead, MAGE encodes expert floorplanning knowledge through structured natural-language directives, visual checks, and iterative refinement criteria.

Core Architecture

MAGE operates through a six-phase workflow orchestrated by ten specialized agents:

1. Data Parsing & Knowledge Extraction: Extracts geometry, hierarchy, and connectivity, querying a knowledge corpus for prior placement patterns and known issues.
2. Hierarchy-Aware Grouping: Organizes macros into groups based on hierarchical prefixes and identifies families of structurally similar groups to share reference placements.
3. Group Placement: Iteratively assigns bounding boxes to macro groups using connectivity and IO-aware placement, guided by visual flyline diagrams to ensure strong connectivity and IO access.
4. Feasibility Checking: Verifies that groups can physically fit within their assigned bounding boxes, recursively partitioning large groups if necessary.
5. Macro Placement: Generates a reference placement for a representative group in each family, adhering to six spatial principles (boundary placement, macro stacking, pin-aware orientation, IO keepouts, pocket minimization, and whitespace quality). This reference is then transformed and merged across the family.
6. Independent Validation: A stateless auditor performs a "second-pass" visual and rule-based audit to catch issues missed by the generation loop, such as blocked pin access or fragmented whitespace.

Multimodal Reasoning and Visual Feedback

A distinguishing feature of MAGE is its mandatory visualization gates. At key stages, the framework renders placement diagrams (e.g., flylines, whitespace heatmaps, boundary coverage maps) which are inspected by Vision-Language Models (VLMs). The agents extract visual-only facts (e.g., "pin edges face into adjacent macros") and generate actionable refinement suggestions. This ensures that spatial properties not easily captured by numerical metrics are explicitly evaluated.

Tournament-Style Refinement

MAGE supports a Go-With-The-Winners (GWTW) tournament mode. Multiple candidate placements are generated in parallel, evaluated using early global routing (eGR) wirelength as a proxy for PPA, and ranked. The top performers are analyzed to identify discrepancies contributing to wirelength, and their insights are used to refine guidance for subsequent rounds. This evolutionary process allows the framework to explore multiple trajectories rather than following a single optimization path.

3. Key Contributions

• Agentic Framework for Human-Like Placement: MAGE is the first agentic macro placement framework explicitly designed to generate human-like solutions by decomposing the task into a structured, multi-phase pipeline rather than a monolithic optimization.
• Principles and Metrics for Human-Likeness: The authors formalize six spatial principles of expert floorplanning and introduce four quantitative metrics to measure "human-likeness":
  • Notch Score: Measures jaggedness in macro silhouettes.
  • Whitespace Score: Evaluates the rectangularity and compactness of free space.
  • Pocket Score: Penalizes small, unusable free-space regions.
  • Alignment Score: Assesses boundary placement, edge alignment, corner occupation, and stacking correctness.
• Two-Stage Synthesis: The framework separates structural organization (Phase 1–5) from implementation metric optimization (Tournament refinement), preserving structural integrity while improving timing and wirelength.
• Generalization: The approach demonstrates zero-shot transfer to unseen designs, rectilinear floorplans, and high-utilization settings without design-specific retraining.

4. Experimental Results

The framework was evaluated on nine designs across NanGate45 (NG45) and GlobalFoundries 12nm (GF12) enablements, comparing against a human expert, two commercial macro placers (Comm-1, Comm-2), and OpenROAD's Hier-RTLMP.

• PPA Improvements:
  • vs. Human Expert: MAGE improved Worst Negative Slack (WNS) by 18.3% and Total Negative Slack (TNS) by 72.5% on NG45 designs, with comparable wirelength and power.
  • vs. Commercial Placers: MAGE achieved geometric-mean improvements of 11.1%–19.3% in WNS and 70.0%–74.0% in TNS over commercial baselines, though with a trade-off of higher routed wirelength (rWL) and power.
  • vs. Hier-RTLMP: MAGE improved WNS by 47.0% and TNS by 80.4%, while also reducing rWL and power.
• Human-Likeness: MAGE outperformed all baselines on the overall human-likeness score ($S_{HM}$) by 6%–48%.
• Case Studies:
  • Unseen Designs: On two GF12 designs not used in training, MAGE improved TNS by 27%–40% over commercial baselines.
  • Rectilinear Floorplans: MAGE improved WNS on dense rectilinear designs (>70% utilization), where commercial tools often produce fragmented placements.
  • High Utilization: Under tight area constraints (up to 80% utilization), MAGE maintained better timing closure than commercial tools by preserving organized routing channels.

5. Significance and Claims

The paper claims that human-like macro organization serves as a useful structural prior for downstream routing and timing closure. The results suggest that the specific structural regularities enforced by MAGE (e.g., continuous whitespace, aligned stacks) directly contribute to better timing outcomes (reduced clock skew and congestion) even when total wirelength is higher.

The authors emphasize that MAGE represents a shift from treating macro placement as a purely numerical optimization problem to a structured design process guided by expert principles. By encoding these principles via multimodal reasoning and visual feedback, the framework can replicate the organizational logic of human experts without requiring expensive training on labeled data. The work demonstrates that agentic, multimodal systems can effectively capture and apply complex physical-design judgment, potentially reducing the manual refinement time that currently dominates industrial flows.