Skip to content

MAGE: Human-Like Macro Placement via Agentic Multimodal Reasoning

🕒 Published (v1): 2026-07-20 22:02 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

MAGE (Macro Placement Agentic Engine) is a multimodal multi-agent framework that decomposes VLSI macro placement refinement into a six-phase, ten-agent pipeline guided by structured natural-language floorplanning rules and visual feedback. Rather than training on labeled placement data, it encodes expert knowledge through prompts, a knowledge corpus, and a tournament-style evolutionary refinement mode. On nine chip designs, it outperforms commercial placers and human experts on both timing metrics and four novel structural "human-likeness" scores.

Problem

Automated macro placement tools produce legal placements but systematically miss the spatial regularity, hierarchical coherence, and structural organization that expert physical-design (PD) engineers apply—resulting in floorplans that require substantial manual refinement before downstream implementation. No prior automated method explicitly targets this structural, human-like quality alongside conventional PPA (power, performance, area) objectives, and no formal metrics exist to quantify it.

Method

MAGE wraps a vision-language model (VLM) inside a multi-agent harness with six sequential phases and ten specialized sub-agents:

  1. Phase 1 — Data Parsing & Knowledge Extraction: A parser agent extracts netlist geometry, hierarchy, connectivity, and IO/pin data; it also queries a design-specific knowledge corpus \(K\) for known floorplan issues.
  2. Phase 2 — Hierarchy-Aware Grouping: A grouping agent forms macro groups \(G = \{g_1,\ldots,g_k\}\) by hierarchy, then clusters them into structurally similar families \(Q_1,\ldots,Q_p\) to enable template reuse.
  3. Phase 3 — Group Placement (iterative): A group-placement agent assigns bounding boxes \(B_g\) to groups using connectivity- and IO-aware logic, iterating until weighted flyline-length reduction falls below 2% or 5 iterations. A mandatory visualization gate fires each iteration: the agent renders a flyline diagram, extracts visual observations, and merges them with automated checks before accepting or revising.
  4. Phase 4 — Feasibility Checking: A feasibility checker verifies each group fits its bounding box (area, packing patterns, routing-channel spacing); groups exceeding size threshold \(\tau_{\max}\) are recursively partitioned.
  5. Phase 5 — Macro Placement: For each family \(Q_j\), a reference group \(g_{\text{ref}}\) receives a placement \(P_{\text{ref}}\) generated under the six structural principles; a validation agent checks legality/pin access/keepouts/whitespace; the placement is then transformed (rotation/reflection/translation) and merged into remaining groups via Algorithm 2.
  6. Phase 6 — Independent Validation + Audit: A stateless independent auditor performs a full chip-level audit (overlap, boundary clearance, channel spacing); failures feed back to Phase 5 or Phase 3 for up to \(N_{\max}=3\) revision iterations.

A seventh optional stage, tournament-style refinement (Go-With-The-Winners, GWTW), spawns \(N_v=6\) variants, runs \(R=6\) rounds keeping \(N_s=3\) survivors per round based on early global-routing (eGR) wirelength from Cadence Innovus, and employs two additional agents (analysis + reflection) to propagate feedback from higher-quality survivors.

Human-likeness is quantified by four metrics normalized to \([0,1]\): - \(S_{\text{notch}} = \max\!\left(0,1 - \tfrac{\sum_i d_i \ell_i}{L_{\text{total}} \cdot 2c_w}\right)\): penalizes jagged macro silhouettes. - \(S_{\text{ws}}\): area-weighted mean rectangularity of free-space regions (compact usable whitespace). - \(S_{\text{pocket}} = 1 - A_{\text{pocket}}/A_{\text{free}}\): penalizes narrow/unusable residual spaces. - \(S_{\text{align}} = 0.3 S_{\text{snap}} + 0.2 S_{\text{edge}} + 0.2 S_{\text{corner}} + 0.3 S_{\text{stack}}\): boundary adherence, inter-macro alignment, corner occupation, and stacking regularity. - \(S_{\text{HM}} = \tfrac{1}{4}(S_{\text{notch}} + S_{\text{ws}} + S_{\text{pocket}} + S_{\text{align}})\).

Key Contributions

  • Multi-agent six-phase placement harness with ten specialized sub-agents handling decomposed sub-tasks (parsing, grouping, placement, feasibility, validation, audit, tournament).
  • Formalization of six human-like floorplanning principles: boundary placement, macro stacking, pin-aware orientation, IO keepouts, pocket minimization, whitespace quality — encoded as natural-language directives rather than learned from data.
  • Four human-likeness metrics (\(S_{\text{notch}}, S_{\text{ws}}, S_{\text{pocket}}, S_{\text{align}}\)) and composite \(S_{\text{HM}}\); first formal quantification of structural floorplan quality.
  • GWTW tournament refinement (multi-variant evolutionary selection guided by eGR wirelength) with analysis and reflection agents propagating cross-iteration feedback.
  • First macro placement evaluation on rectilinear floorplans; evaluation across NanGate45 and GlobalFoundries 12nm, including unseen designs not used in prompt or corpus development.

Results

  • vs. Hier-RTLMP (NanGate45, 3 designs): +47.0% WNS, +80.4% TNS, comparable wirelength and power; \(S_{\text{HM}}\) improvement 6%–48%.
  • vs. human expert (NanGate45, 3 designs): +18.3% WNS, +72.5% TNS.
  • vs. commercial macro placer (9 designs, NG45 + GF12): geometric-mean +11.1%–19.3% WNS, +70.0%–74.0% TNS; \(S_{\text{HM}}\) +6%–48%.
  • Dense rectilinear designs (>70% utilization): +5%–40% WNS, up to +74% TNS vs. commercial baseline.
  • Unseen designs (2 designs, not used in prompt/corpus development): +6%–27% WNS, +26%–40% \(S_{\text{HM}}\) vs. commercial baseline.

Limitations

  • Requires an initial placement as input (refinement framework, not placement from scratch).
  • eGR wirelength (Cadence Innovus) serves as the GWTW objective proxy; full sign-off PPA requires additional downstream steps.
  • Tournament mode multiplies VLM inference cost by up to \(N_v \times R = 36\) placements evaluated; no wall-clock runtime reported.
  • Human-likeness metrics were calibrated against a specific set of expert-designed placements and may not generalize to all design styles or technology nodes.
  • Knowledge corpus construction is design-family-specific; its curation cost is not quantified.
  • The six structural principles and their encoding as prompts are not exhaustive; the authors note additional expert guidance can extend them.
  • Evaluation of human-expert baseline is limited to three NanGate45 designs; no GF12 human-expert comparison is available.

Relevance to Harnesses / Meta-Harnesses

MAGE is a concrete instantiation of a domain-specific multi-agent harness: it orchestrates ten specialized sub-agents through a stateful six-phase pipeline with conditional back-edges (Phase 6 → Phase 5 → Phase 3), mandatory visualization gates, and a separate evolutionary meta-loop (GWTW tournament). This architecture directly exemplifies the harness pattern of wrapping foundation-model calls in structured control flow — phase gating, iteration budgets, and validation checkpoints — rather than issuing a single monolithic prompt. The tournament refinement layer functions as a meta-harness layer above the core pipeline, spawning multiple independent pipeline runs and using a reflection agent to synthesize cross-run learning, analogous to outer-loop orchestrators in general agentic frameworks. For researchers tracking harness design, MAGE offers a case study where domain expert knowledge is externalized into a reusable knowledge corpus and natural-language rule system rather than baked into model weights, enabling zero-shot transfer to unseen designs — a key generalizability property sought in general-purpose meta-harnesses.