Pith. sign in

REVIEW 4 major objections 5 minor 50 references

MAGE claims that encoding expert floorplanning heuristics as natural-language rules and visual checks, in a ten-agent pipeline, produces human-like macro placements that improve post-route timing by 11–19% in WNS and 70–74% in TNS over comm

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A multi-agent vision-language framework for chip macro placement improves post-route timing (WNS/TNS) over commercial, human, and Hier-RTLMP baselines while trading higher wirelength, and introduces four human-likeness metrics.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A serious, honest integration of agentic LLM reasoning into macro placement with an unproven tournament mechanism; send to reviewers, but the eGR proxy and single-run results need to be addressed. the 4 major comments →

arxiv 2607.18536 v1 pith:L2JDBQF5 submitted 2026-07-20 cs.AI cs.RO

MAGE: Human-Like Macro Placement via Agentic Multimodal Reasoning

Andrew B. Kahng , Sayak Kundu , Bodhisatta Pramanik This is my paper
classification cs.AI cs.RO
keywords macro placementchip floorplanningmulti-agent systemsvision-language modelmultimodal reasoningtiming closurerouted wirelengthhuman-likeness metrics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

MAGE tackles macro placement, the task of placing large pre-designed blocks in a chip floorplan. The paper's premise is that expert engineers arrange macros according to structural principles—boundary placement, regular stacking, IO keepouts, avoiding unusable pockets, and preserving contiguous whitespace—and that these principles, not raw wirelength optimization, are what make a floorplan easy to route and close timing. Rather than training a model on labeled placements, MAGE encodes those principles as natural-language directives and visual validation gates in a ten-agent pipeline, then uses a tournament to refine candidate placements. Across nine designs and two technology nodes, it reports large timing improvements over commercial and hierarchy-aware baselines, at the cost of higher routed wirelength in most comparisons. The decisive assertion, supported by case studies on unseen, rectilinear, and high-utilization designs, is that human-like macro organization is a useful structural prior for downstream routability and timing closure.

Core claim

The paper's central discovery is that human-like macro organization is a useful structural prior for downstream timing closure. It shows that a purely prompt-and-vision-driven system, with no training on labeled placement data, can produce macro placements that align with six expert principles and, across nine benchmarks, improve geometric-mean worst negative slack by 11.1%–19.3% and total negative slack by 70.0%–74.0% over commercial macro placers. On the three designs with human-expert baselines, it improves WNS by 18.3% and TNS by 72.5% while keeping wirelength and power comparable. The paper attributes the gains to lower clock insertion delay, reduced clock skew, and lower congestion aro

What carries the argument

The framework is MAGE, a multi-agent system of ten specialized agents organized into six phases: data parsing, hierarchy-aware grouping, group placement, feasibility checking, macro reference placement, and independent validation. Two mechanisms carry the argument: mandatory visual gates, in which a vision-language model inspects rendered flyline diagrams, whitespace heatmaps, and boundary-coverage maps and must report visual-only facts; and a go-with-the-winners tournament that runs six placement variants in parallel, ranks them by a fast pre-routing wirelength estimate, keeps the top three survivors, and propagates analysis and reflection feedback. The four human-likeness metrics—notch, wh

Load-bearing premise

The load-bearing premise is that the fast pre-routing wirelength estimate used to rank candidate placements in the tournament faithfully predicts the final post-route timing metrics (WNS and TNS) that the headline gains are measured on; the paper asserts this correlation but does not show it.

What would settle it

Run one tournament round on a design and record, for every variant, both the pre-routing wirelength that selects survivors and the final post-route WNS/TNS. If the variant with the best proxy is not generally the best-timing variant, or if the rank correlation between the proxy and timing is near zero or negative, then the tournament's contribution to the reported timing improvements is unsupported.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Macro placement can be framed as a knowledge-encoding problem: expert floorplanning rules written in natural language, plus visual checks, can substitute for labeled training data.
  • Structural regularity—boundary alignment, regular stacks, and contiguous whitespace—is worth pursuing even when it increases total routed wirelength, because the timing gains come from reduced clock skew and congestion.
  • The tournament's roughly 9% wirelength improvement from analysis and reflection feedback suggests that iterative, feedback-driven refinement is where much of the placement-quality gain lives.
  • Zero-shot transfer to unseen, rectilinear, and high-utilization designs implies that the expert principles, rather than memorized design-specific layouts, are doing the work.
  • The four human-likeness metrics give designers a vocabulary for floorplan quality beyond PPA and could be used as design-time checks for structural regularity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A direct test of the paper's mechanism would be to run the tournament with WNS or TNS (or a fast timing estimate) as the in-loop objective; if the reported timing gains persist, the structural-prior story is confirmed, and if they vanish, the gains likely come from selection rather than the human-like rules.
  • The frequent tradeoff between higher total wirelength and better timing suggests that total wirelength is the wrong proxy for the final objective; a natural extension is to optimize congestion-aware or critical-path-aware wirelength inside the tournament.
  • Because the human-likeness metrics are computable from a floorplan alone, they could be inserted as regularizers into learned placers or used to build a fast surrogate for routing outcomes—something the paper does not propose.
  • The finding that anonymizing design names degrades timing hints that semantic labels encode regular arrays and symmetry that connectivity alone does not; an inference worth testing is whether graph-isomorphism-style symmetry detection could recover this without human-readable names.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes MAGE, a multimodal multi-agent framework for macro placement refinement. Ten agents execute a six-phase pipeline (parsing, hierarchy-aware grouping, group placement, feasibility, reference placement/transform-and-merge, independent validation) and optionally a Go-With-The-Winners tournament that selects among variants by Innovus early-global-routing (eGR) routed wirelength. Expert floorplanning knowledge is encoded as natural-language directives, visual checks, and validation criteria rather than learned from labeled placements. Four human-likeness metrics are introduced. Experiments on nine designs across NanGate45 and GF12 report geometric-mean WNS/TNS improvements of 11.1-19.3%/70-74% over two commercial macro placers, 18.3%/72.5% over a human expert, and 47.0%/80.4% over Hier-RTLMP, at the cost of higher wirelength and modestly higher power versus the commercial placers. Ablations show the knowledge corpus and visual feedback affect timing; additional case studies address anonymized netlists, unseen designs, rectilinear floorplans, and high utilization. The paper's central claim is that human-like macro organization provides a useful structural prior for downstream timing closure.

Significance. The result is significant if it holds: it would demonstrate that a multimodal, multi-agent LLM system can encode expert physical-design judgment and improve post-route timing over commercial and human baselines without any training on labeled placements. The evaluation has real strengths: it uses a standard Innovus post-route flow, reports full PPA across two enablements, includes a human-expert baseline, presents ablations and cost/runtime data, and points to a public repository. However, two caveats are load-bearing: the tournament's selection metric (eGR routed wirelength) is asserted to be a strong proxy for downstream PPA but the correlation is never shown, and all results are single stochastic runs with no error bars. These caveats currently prevent full confidence in the causal attribution claimed in the abstract and conclusion.

major comments (4)
  1. [§IV-G, Algorithm 3 (lines 9-10, 23); §VI-A] The tournament selects P* by minimum eGR routed wirelength, and Section IV-G states that eGR rWL is 'empirically a strong proxy for downstream PPA metrics (Table II)'. Table II, however, contains only post-route rWL, power, WNS, and TNS; no eGR rWL values and no eGR-versus-post-route correlation are provided anywhere. Since Section VI-A explicitly distinguishes post-route rWL from eGR rWL, the proxy claim is not tautological. All headline WNS/TNS improvements are for the tournament-selected P*, so if eGR rWL correlates weakly or negatively with timing, the reported gains could come from the base six-phase pipeline or from best-of-N selection rather than from the mechanism the paper emphasizes. Please supply the eGR-vs-post-route correlation per benchmark, or an ablation that selects on post-route timing.
  2. [§VI experimental setup; Table II] The 'Human' baseline is not described. The reader is told only that human-expert baselines are available on the three NG45 designs, with no information about who produced the placements, under what instructions, whether one or multiple experts were involved, or whether the placements come from a published benchmark. Since the abstract and Section VI-A claim an 18.3% WNS / 72.5% TNS improvement over the human expert, this baseline must be defined and, ideally, verified or cited precisely.
  3. [Tables II, V-VIII, X; Figures 8 and 12] All reported results appear to be single runs. The pipeline is stochastic (LLM sampling, tournament variants, validation iterations), yet no standard deviations, number of seeds, or repetition protocol are reported. The per-run cost is acknowledged, but without variance information the geometric-mean improvements cannot be distinguished from selection effects among the sampled variants. At minimum, seeded reruns for one NG45 and one GF12 design, and confidence intervals for the headline metrics, would be needed to support the precision of the claimed improvements.
  4. [§III, §III-A, §IV] The four human-likeness metrics are direct operationalizations of the six principles that are also encoded as prompts, validation checks, and visual criteria in the MAGE pipeline. Consequently, Table IV and Figure 8 partly measure the method against its own rule system. The paper's 'Metric design and limitations' paragraph is candid about this, but the central premise that 'human-like macro organization provides a useful structural prior' would be materially stronger with an external validation, e.g., independent human judgments of placement quality or a comparison against human placements not used to define the metrics.
minor comments (5)
  1. [§VI-D.1, Table VII] The anonymization study is confounded: anonymizing names also disables the knowledge corpus K, so the observed timing degradation could be due to K removal rather than to loss of semantic names. The text claims the study isolates what names contribute beyond structure, but K removal is a second independent change. Please ablate name anonymization with K retained, or explicitly acknowledge the confound.
  2. [Figure 8] The Pearson correlation is reported as r=+0.38, p=0.053, which is not statistically significant at the 0.05 level; the text should not rely on it without reporting the sample size and should give more weight to the Spearman result or describe the relationship as suggestive.
  3. [§IV-D, Eq. (1)] The notation for effective column widths and row heights (\hat{w}_i, \hat{h}_j) and channel spacings (c_x, c_y) is introduced only in prose; a one-sentence definition in the equation caption or immediately before would improve readability.
  4. [§VI-D.3, Table IX] For the Mixed rectilinear variant, MAGE does not always improve TNS (Ariane-Mixed and BP-Mixed show roughly equal or worse TNS). The text says only that WNS improves on all four variants; the TNS tradeoff should be stated explicitly for these cases.
  5. [Figure 12 caption] The caption says 'Normalized Routed Wirelength', which could be confused with the post-route rWL reported in Table II. Since the tournament uses eGR rWL, the caption and axis label should say 'eGR routed wirelength'.

Circularity Check

1 steps flagged

Human-likeness metrics are defined from the same principles MAGE is built to enforce; PPA claims remain externally grounded.

specific steps
  1. self definitional [Section III-A (metrics), Section IV-A (pipeline rules), Section VI-B (Table IV)]
    "These metrics translate the placement principles in Section III into measurable scores. ... MAGE encodes them as shared rules used throughout the placement pipeline to maintain consistent floorplanning intent."

    The overall human-likeness score SHM is the equal-weighted mean of four metrics that are explicitly defined as measurable translations of the six placement principles. The same principles are then encoded into MAGE's prompts, validation scripts, visual gates, and independent auditor, and are used to generate and accept placements. Thus Table IV's reported 6%-48% SHM improvement is substantially a check of MAGE against its own encoded rulebook: the high human-likeness score is built into the generator/metric construction rather than discovered from an independent source. This circularity is confined to the human-likeness evaluation; WNS/TNS/rWL come from Cadence Innovus, and tournament selection optimizes eGR routed wirelength rather than SHM, so the PPA claims retain independent content.

full rationale

The paper's headline PPA claims are not circular: WNS, TNS, routed wirelength, and power are all measured by Cadence Innovus after a standard place-and-route flow, and the tournament (Algorithm 3) selects P* by minimum eGR routed wirelength, not by the paper's human-likeness metrics. The asserted proxy relationship between eGR rWL and downstream PPA is not validated in Table II, but that is an unsupported empirical claim, not a definitional reduction, so it does not constitute circularity under the given criteria. The main circularity concern is the human-likeness evaluation: the four metrics are defined from the six principles, and those same principles are encoded into MAGE's prompts, validation checks, and visual gates. Consequently, reporting MAGE as 6%-48% more human-like on those metrics is partly measuring the method against its own rulebook. The comparison against the human-expert baseline and the external Innovus-based PPA numbers provide some independent anchor, so the circularity is partial and confined to a supporting claim rather than to the central timing results. Self-citations such as [9], [10] for the evaluation flow are methodology citations, not load-bearing uniqueness or ansatz justifications, and no self-citation chain forces the main result.

Axiom & Free-Parameter Ledger

9 free parameters · 6 axioms · 0 invented entities

The central claims rest on hand-set parameters (channel widths, weights, tournament sizes), an unverified VLM reliability assumption, an unquantified eGR-to-PPA proxy, and the paper's own human-likeness formalization. No new physical entities are introduced; the 'agents', 'knowledge corpus', and metrics are software constructs.

free parameters (9)
  • channel width c_w = 20 um (NG45), 10 um (GF12)
    Hand-set in Section III-A footnote 5; used in all human-likeness metrics and feasibility checks; not tuned per method, but choice affects scores and may favor MAGE's boundary-stacked style.
  • IO keepout multiplier = 10*c_w
    Footnote 5 sets keepout to 10*c_w; used by validation and feasibility; affects placement legality and IO clearance.
  • alignment score weights (w1..w4) = 0.3, 0.2, 0.2, 0.3
    Section III-A defines S_align with fixed weights; no sensitivity analysis; directly defines the metric on which MAGE claims superiority.
  • significant free-space threshold = 4*c_w^2
    Section III-A; used to define pockets and whitespace score thresholds.
  • alignment tolerance tau = c_w/2
    Section III-A; coordinate-level comparisons for edge alignment, corner occupation, and stacking validity.
  • tournament defaults (Nv, Ns, R) = 6, 3, 6
    Section IV-G and Table I; hand-set defaults; no ablation of these values (paper notes R=3 could suffice).
  • validation iteration limit Nmax = 3
    Table I; pipeline convergence limit.
  • group-size threshold tau_max = 30
    Phase 4 footnote 12; recursive partitioning threshold.
  • group placement loop termination = 5 iterations or <2% flyline reduction
    Phase 3; hand-set termination criteria affecting group-box convergence.
axioms (6)
  • domain assumption Claude Opus 4.6 handles spatial/visual reasoning from rendered floorplan images and natural-language directives reliably.
    All ten agents depend on the VLM; ablation in Section VI-C (2 designs) is the only evidence, with no systematic error analysis.
  • domain assumption Innovus eGR routed wirelength is a strong proxy for downstream post-route PPA.
    Stated in Section IV-G without a correlation table; tournament selection uses eGR, not WNS/TNS.
  • ad hoc to paper The six spatial principles in Section III characterize human-like macro placement and are beneficial for timing.
    These principles are the paper's own formalization; used both to prompt MAGE and to define evaluation metrics; support from Figure 8 is weak (r=0.38, p=0.053).
  • domain assumption Commercial baselines Comm-1/Comm-2 and the Human baseline are representative and correctly run.
    Baselines are unnamed or proprietary, and Human baseline provenance is unspecified; comparisons cannot be independently replicated.
  • domain assumption The obfuscation step prevents the LLM from using tool/enablement-specific information.
    Section VI states names are anonymized and dimensions scaled, but no test of information leakage is provided.
  • domain assumption Post-route flow and reporting are identical across methods, so all PPA differences come from macro placement.
    Stated in Section VI-A; relies on correct execution of the same Innovus script for all placements.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of MAGE: Human-Like Macro Placement via Agentic Multimodal Reasoning." pith.science (2026). https://pith.science/paper/L2JDBQF5

@misc{pith2026260718536,
  author       = {Pith},
  title        = {Pith review of: MAGE: Human-Like Macro Placement via Agentic Multimodal Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L2JDBQF5}},
  note         = {Machine review of arXiv:2607.18536}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Macro placement still requires substantial manual refinement in industrial physical design flows. We present MAGE (Macro Placement Agentic Engine), a multimodal multi-agent framework for macro placement refinement. MAGE decomposes the macro placement task into a six-phase workflow that combines structured floorplanning rules, visual checks, and iterative refinement. Expert floorplanning knowledge is encoded through natural-language directives and validation criteria, rather than learned from labeled placement data. A tournament-style refinement mode evaluates multiple candidate placements and propagates feedback from higher-quality solutions. We also introduce four metrics for quantifying human-likeness in macro placement: notch score, whitespace score, pocket score, and alignment score. These metrics capture structural properties used by expert designers but not directly measured by conventional PPA metrics. Across nine designs in NanGate45 and GlobalFoundries 12nm enablements, MAGE achieves geometric-mean improvements of 11.1%-19.3% in WNS and 70.0%-74.0% in TNS over commercial macro placers. On the three NanGate45 designs, for which human-expert and Hier-RTLMP baselines are available, MAGE improves WNS and TNS by 18.3% and 72.5% over the human expert, and by 47.0% and 80.4% over Hier-RTLMP, with comparable wirelength and power. On human-likeness metrics, MAGE improves the overall score by 6%-48% over all baselines. Additional case studies on anonymized netlists, unseen designs, dense rectilinear floorplans, and high-utilization settings show that the framework transfers to new placement settings without design-specific retraining.

Figures

Figures reproduced from arXiv: 2607.18536 by Andrew B. Kahng, Bodhisatta Pramanik, Sayak Kundu.

Figure 1
Figure 1. Figure 1: Six spatial principles for human-like macro placement [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of MAGE framework. “decompose–place–validate–refine” workflow, in which place￾ment decisions are guided by design context and visual feed￾back. This workflow is implemented through ten sub-agents ( [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Progressive refinement of a macro placement through [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Some visual artifacts used for feedback in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Excerpt of a visual feedback report generated by the [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Evaluation flow for macro placements produced by [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Macro placement comparisons across NG45 and GF12 [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Per-design min-max-normalized TNS versus human￾likeness score SHM across all (design, method) pairs. For each design, TNS is min-max normalized across methods so that 1 denotes the best TNS and 0 the worst. Higher human-likeness is positively associated with better post-route timing (Pearson r = +0.38, p = 0.053; Spearman ρ = +0.42, p = 0.030). B. Evaluation of Human-Likeness Table IV summarizes the geomet… view at source ↗
Figure 9
Figure 9. Figure 9: Effect of visual feedback on Ariane and BP (NG45). [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visual comparison of Comm-1 and MAGE on two unseen designs. Top: VTA (GF12). Bottom: GENESYS01 (GF12). Left: Comm-1. Right: MAGE. higher rWL [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Placement comparison on dense rectilinear floorplans [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Normalized routed wirelength across tournament [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 4 linked inside Pith

  1. [1]

    Fixed-outline floorplanning: Enabling hierarchical design

    S. N. Adya and I. L. Markov, “Fixed-outline floorplanning: Enabling hierarchical design”,IEEE TVLSI, 11(6) (2003), pp. 1120–1135

    2003

  2. [2]

    ‘Go with the winners’ algorithms

    D. Aldous and U. Vazirani, “‘Go with the winners’ algorithms”,Proc. IEEE FOCS, 1994, pp. 492–501

    1994

  3. [3]

    AutoDMP: Automated DREAMPlace-based macro placement

    A. Agnesina, P. Rajvanshi, T. Yang, G. Pradipta, A. Jiao, B. Keller et al., “AutoDMP: Automated DREAMPlace-based macro placement”,Proc. ISPD, 2023, pp. 149–157

    2023

  4. [4]

    A highly configurable hardware/software stack for DNN inference acceleration

    S. Banerjee, S. Burns, P. Cocchini, A. Davare, S. Jain, D. Kirkpatrick et al., “A highly configurable hardware/software stack for DNN inference acceleration”,arXiv:2111.15024, 2021

    Pith/arXiv arXiv 2021

  5. [5]

    B*-trees: A new representation for non-slicing floorplans

    Y .-C. Chang, Y .-W. Chang, G.-M. Wu, and S.-W. Wu, “B*-trees: A new representation for non-slicing floorplans”,Proc. DAC, 2000, pp. 458– 463

    2000

  6. [6]

    A new multilevel framework for large- scale interconnect-driven floorplanning

    T. Chen, Y . Chang, and S. Lin, “A new multilevel framework for large- scale interconnect-driven floorplanning”,IEEE TCAD, 27(2) (2008), pp. 286–294

    2008

  7. [7]

    Routability-driven blockage-aware macro placement

    Y .-F. Chen, C.-C. Huang, C.-H. Chiou, Y .-W. Chang and C.-J. Wang, “Routability-driven blockage-aware macro placement”,Proc. DAC, 2014, pp. 1–6

    2014

  8. [8]

    RePlAce: Advancing solution quality and routability validation in global placement

    C.-K. Cheng, A. B. Kahng, I. Kang and L. Wang, “RePlAce: Advancing solution quality and routability validation in global placement”,IEEE TCAD, 38(9) (2018), pp. 1717–1730

    2018

  9. [9]

    As- sessment of reinforcement learning for macro placement

    C.-K. Cheng, A. B. Kahng, S. Kundu, Y . Wang and Z. Wang, “As- sessment of reinforcement learning for macro placement”,Proc. ISPD, 2023, pp. 158–166

    2023

  10. [10]

    An updated assessment of reinforcement learning for macro placement

    C.-K. Cheng, A. B. Kahng, S. Kundu, Y . Wang and Z. Wang, “An updated assessment of reinforcement learning for macro placement”, IEEE TCAD, (2025) (DOI 10.1109/TCAD.2025.3644293)

    arXiv 2025

  11. [11]

    Circular- contour-based obstacle-aware macro placement

    C.-H. Chiou, C.-H. Chang, S.-T. Chen and Y .-W. Chang, “Circular- contour-based obstacle-aware macro placement”,Proc. ASP-DAC, 2016, pp. 172–177

    2016

  12. [12]

    Hierarchical global floorplacement using simulated annealing and network flow area migration

    W. Choi and K. Bazargan, “Hierarchical global floorplacement using simulated annealing and network flow area migration”,Proc. DATE, 2003, pp. 1104–1105

    2003

  13. [13]

    Design-hierarchy aware mixed-size placement for routability optimization

    Y .-L. Chuang, G.-J. Nam, C. J. Alpert, Y .-W. Chang, J. Roy and N. Viswanathan, “Design-hierarchy aware mixed-size placement for routability optimization”,Proc. ICCAD, 2010, pp. 663–668

    2010

  14. [14]

    Fast floorplanning by lookahead enabled recursive bipartitioning

    J. Cong, M. Romesis, and J. R. Shinnerl, “Fast floorplanning by lookahead enabled recursive bipartitioning”,IEEE TCAD, 25(9) (2006), pp. 1719–1732

    2006

  15. [15]

    Profile-guided microarchitectural floorplanning for deep submicron processor design

    M. Ekpanyapong, J. Minz, T. Watewai, H.-H. S. Lee and S. K. Lim, “Profile-guided microarchitectural floorplanning for deep submicron processor design”,IEEE TCAD, 25(7) (2006), pp. 1289–1300

    2006

  16. [16]

    VeriGOOD-ML: An open-source flow for automated ML hardware synthesis

    H. Esmaeilzadeh, S. Ghodrati, J. Gu, S. Guo, A. B. Kahng, J. K. Kim et al., “VeriGOOD-ML: An open-source flow for automated ML hardware synthesis”,Proc. ICCAD, 2021, pp. 1–7

    2021

  17. [17]

    Automated QoR improve- ment in OpenROAD with coding agents

    A. Ghose, J. Jang, A. B. Kahng and J. Lee, “Automated QoR improve- ment in OpenROAD with coding agents”,arXiv:2601.06268, 2026

    arXiv 2026

  18. [18]

    A GA with heuristic-based decoder for IC floorplanning

    B. H. Gwee and M. H. Lim, “A GA with heuristic-based decoder for IC floorplanning”,Integration, 28(2) (1999), pp. 157–172

    1999

  19. [19]

    Learn to floorplan through acquisition of effective local search heuristics

    Z. He, Y . Ma, L. Zhang, and X. Zhou, “Learn to floorplan through acquisition of effective local search heuristics”,Proc. ICCD, 2020, pp. 324–331

    2020

  20. [20]

    NTUplace4h: A novel routability-driven placement algorithm for hierarchical mixed-size circuit designs

    M.-K. Hsu, Y .-F. Chen, C.-C. Huang, S. Chou, T.-H. Lin, T.-C. Chen et al., “NTUplace4h: A novel routability-driven placement algorithm for hierarchical mixed-size circuit designs”,IEEE TCAD, 33(12) (2014), pp. 1914–1927

    2014

  21. [21]

    Fast multilevel floorplanning for large scale modules

    C.-C. Hu, D.-S. Chen, and Y .-W. Wang, “Fast multilevel floorplanning for large scale modules”,Proc. ISCAS, 2004, pp. 205–208

    2004

  22. [22]

    RTL-MP: Toward practical, human-quality chip planning and macro placement

    A. B. Kahng, R. Varadarajan, and Z. Wang, “RTL-MP: Toward practical, human-quality chip planning and macro placement”,Proc. ISPD, 2022, pp. 3–11

    2022

  23. [23]

    Hier-RTLMP: A hierar- chical automatic macro placer for large-scale complex IP blocks

    A. B. Kahng, R. Varadarajan, and Z. Wang, “Hier-RTLMP: A hierar- chical automatic macro placer for large-scale complex IP blocks”,IEEE TCAD, 43(5) (2024), pp. 1552–1565

    2024

  24. [24]

    MAPLE: Multilevel adaptive placement for mixed-size designs

    M.-C. Kim, N. Viswanathan, C. J. Alpert, I. L. Markov and S. Ramji, “MAPLE: Multilevel adaptive placement for mixed-size designs”,Proc. ISPD, 2012, pp. 193–200

    2012

  25. [25]

    Bus-aware microarchitectural floorplanning

    D. H. Kim and S. K. Lim, “Bus-aware microarchitectural floorplanning”, Proc. ASP-DAC, 2008, pp. 204–208

    2008

  26. [26]

    Optimization by simulated annealing

    S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi, “Optimization by simulated annealing”,Science, 220(4598) (1983), pp. 671–680

    1983

  27. [27]

    Regularity-aware routability-driven macro placement methodology for mixed-size circuits with obstacles

    J.-M. Lin, Y .-L. Deng, S.-T. Li, B.-H. Yu, L.-Y . Chang and T.-W. Peng, “Regularity-aware routability-driven macro placement methodology for mixed-size circuits with obstacles”,IEEE TVLSI, 27(1) (2019), pp. 57– 68

    2019

  28. [28]

    Dataflow- aware macro placement based on simulated evolution algorithm for mixed-size designs

    J.-M. Lin, Y .-L. Deng, Y .-C. Yang, J.-J. Chen and P.-C. Lu, “Dataflow- aware macro placement based on simulated evolution algorithm for mixed-size designs”,IEEE TVLSI, 29(5) (2021), pp. 973–984

    2021

  29. [29]

    ePlace-MS: Electrostatics-based placement for mixed-size circuits

    J. Lu, H. Zhuang, P. Chen, H. Chang, C.-C. Chang, Y .-C. Wong et al., “ePlace-MS: Electrostatics-based placement for mixed-size circuits”, IEEE TCAD, 34(5) (2015), pp. 685–698

    2015

  30. [30]

    Chip placement with deep reinforcement learning

    A. Mirhoseini, A. Goldie, M. Yazgan, J. Jiang, E. Songhori, S. Wang et al., “Chip placement with deep reinforcement learning”, arXiv:2004.10746, 2020

    Pith/arXiv arXiv 2004

  31. [31]

    A graph placement methodology for fast chip design

    A. Mirhoseini, A. Goldie, M. Yazgan, J. W. Jiang, E. Songhori, S. Wang et al., “A graph placement methodology for fast chip design”,Nature, 594 (2021), pp. 207–212

    2021

  32. [32]

    VLSI module placement based on rectangle-packing by the sequence-pair

    H. Murata, K. Fujiyoshi, S. Nakatake and Y . Kajitani, “VLSI module placement based on rectangle-packing by the sequence-pair”,IEEE TCAD, 15(12) (1996), pp. 1518–1524

    1996

  33. [33]

    Microarchitecture-aware floorplanning using a statistical design of experiments approach

    V . Nookala, Y . Chen, D. J. Lilja and S. S. Sapatnekar, “Microarchitecture-aware floorplanning using a statistical design of experiments approach”,Proc. DAC, 2005, pp. 579–584

    2005

  34. [34]

    DeFer: Deferred decision making enabled fixed- outline floorplanner

    J. Z. Yan and C. Chu, “DeFer: Deferred decision making enabled fixed- outline floorplanner”,Proc. DAC, 2008, pp. 161–166

    2008

  35. [35]

    An effective floorplan-guided placement algorithm for large-scale mixed-size design

    J. Z. Yan, N. Viswanathan, and C. Chu, “An effective floorplan-guided placement algorithm for large-scale mixed-size design”,ACM TODAES, 19(3) (2014), pp. 1–25

    2014

  36. [36]

    A novel damped-wave framework for macro placement

    C.-H. Chang, Y .-W. Chang, and T.-C. Chen, “A novel damped-wave framework for macro placement”,Proc. ICCAD, 2017, pp. 504–511. 16

    2017

  37. [37]

    ChipNeMo: Domain-adapted LLMs for chip design

    M. Liu, T.-D. Ene, R. Kirby, C. Cheng, N. Pinckney, R. Liang et al., “ChipNeMo: Domain-adapted LLMs for chip design”, arXiv:2311.00176, 2023

    Pith/arXiv arXiv 2023

  38. [38]

    ChatEDA: A large language model powered autonomous agent for EDA

    Z. He, H. Wu, X. Zhang, X. Yao, S. Zheng, H. Zheng, et al., “ChatEDA: A large language model powered autonomous agent for EDA”,IEEE TCAD, 43(10) (2024), pp. 3184–3197

    2024

  39. [39]

    VeriGen: A large language model for Verilog code generation

    S. Thakur, B. Ahmad, H. Pearce, B. Tan, B. Dolan-Gavitt, R. Karri et al., “VeriGen: A large language model for Verilog code generation”, ACM TODAES, 29(3) (2024), pp. 1–31

    2024

  40. [40]

    RTL-aware dataflow-driven macro placement

    A. Vidal-Obiols, J. Cortadella, J. Petit, M. Galceran-Oms and F. Mar- torell, “RTL-aware dataflow-driven macro placement”,Proc. DATE, 2019, pp. 186–191

    2019

  41. [41]

    See it to place it: Evolving macro placements with vision-language models

    I. Uchendu, S. Goel, K. Hou, E. Songhori, K.-H. Lee, J. W. Jiang et al., “See it to place it: Evolving macro placements with vision-language models”,arXiv:2603.28733, 2026

    Pith/arXiv arXiv 2026

  42. [42]

    Accessed: March 11, 2026

    Ariane RISC-V CPU repo. Accessed: March 11, 2026. [Online]. Available: https://github.com/openhwgroup/cva6

    2026

  43. [43]

    Accessed: March 11, 2026

    BlackParrot repo. Accessed: March 11, 2026. [Online]. Available: https://github.com/black-parrot/black-parrot

    2026

  44. [44]

    Accessed: March 11, 2026

    Anthropic’s Claude Opus 4.6. Accessed: March 11, 2026. [Online]. Available: https://www.anthropic.com/news/claude-opus-4-6

    2026

  45. [45]

    Accessed: March 11, 2026

    Hier-RTLMP repo. Accessed: March 11, 2026. [Online]. Available: https://github.com/The-OpenROAD-Project/OpenROAD/tree/master/ src/mpl

    2026

  46. [46]

    https://cadence.com

    Cadence Innovus v21.1. https://cadence.com

  47. [47]

    Accessed: March 11, 2026

    MacroPlacement repo. Accessed: March 11, 2026. [Online]. Available: https://github.com/TILOS-AI-Institute/MacroPlacement

    2026

  48. [48]

    Accessed: March 11, 2026

    MemPool repo. Accessed: March 11, 2026. [Online]. Available: https://github.com/pulp-platform/mempool

    2026

  49. [49]

    Accessed: March 11, 2026

    NanGate45 PDK. Accessed: March 11, 2026. [Online]. Available: https://eda.ncsu.edu/freepdk/freepdk45/

    2026

  50. [50]

    [Online]

    MAGE repository. [Online]. Available: https://github.com/ABKGroup/MAGE Andrew B. Kahngis Distinguished Professor of CSE and ECE at the University of California, San Diego. His interests include IC physical design, the design-manufacturing interface, combinatorial opti- mization, and AI/ML for EDA and IC design. He received the Ph.D. degree in Computer Sci...

    2017

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.