acceptodds
Under review as a conference paper at ICLR 2027

MacroSPA: Spatial Prior Adaptation for Diffusion-based Macro Placement

Abstract

Macro placement determines the spatial organization of large circuit blocks and affects subsequent standard-cell placement and routing. However, the spatial preferences inherited by pretrained placers may not align with target design objectives. We propose **MacroSPA**, a post-training framework for **S**patial **P**rior **A**daptation (SPA) that enables the same pretrained diffusion-based macro placer to acquire distinct spatial priors directly from target context-placement samples. These samples can encode prescribed preferences through synthetic generation or capture layout patterns produced by an existing placement optimizer. SPA adjusts how features of the current layout enter graph message passing and feature fusion while updating only 3.03% of the adapted model's parameters and keeping the pretrained backbone frozen. The framework complements SPA with Minkowski Guided Refinement to reduce residual geometric conflicts and with size-adaptive constraint evaluation for efficient inference on large circuits. Experiments show that MacroSPA acquires distinct spatial priors and can improve downstream outcomes relative to the matched pipeline without SPA. It also achieves competitive routing and post-route performance among the evaluated methods. The relative benefits of different spatial priors vary across circuits and objectives, highlighting the practical value of adapting the inherited spatial prior to the target design.

Primary area:
generative models
Keywords:
Diffusion models, Macro placement, Chip design, EDA

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_JJLJOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MacroSPA adapts a pretrained diffusion macro placer to spatial preferences specified by synthetic or optimizer-generated layouts. It adds Minkowski Guided Refinement and sparse constraint evaluation. Experiments on IBM, ISPD and ChiPBench demonstrate prior adaptation, downstream improvements over the unadapted pipeline, and improved inference scalability.

Critical strengths
  1. SPA adapts a frozen placer with matched controls, showing changes in spatial priors and downstream benefits.
  2. MGR and sparse constraint evaluation provide useful geometric repair and substantial inference savings on large circuits.
Critical weaknesses
  1. Data quality and total cost are insufficiently established. Appendix B.3 samples legal layouts and rewires connectivity to favor short connections; this does not establish quality for realistic fixed netlists. Quality criteria and sensitivity to poorer pretraining/target samples are missing. Dataset-generation costs, including synthetic data and DREAMPlace teacher runs, are unreported. Runtime comparisons with DREAMPlace and other solver-based methods are absent, leaving the practical trade-off unclear.
  2. SPA combines standard residual adapters, LoRA, and head tuning, so its distinct contribution needs stronger evidence. Downstream component ablations and comparisons with parameter-matched adaptation alternatives are missing.
Questions for authors

Please explain the weaknesses above.

6 comments
AuthorsOpenReview ↗

Response to Reviewer JJLJ – Part 1 / 3

Dear Reviewer JJLJ,

Thank you for the detailed and concrete comments. The revision adds a parameter-matched downstream ablation, sensitivity analyses for pretraining data and target samples, cost measurements, and a runtime comparison with DREAMPlace. We respond to each point below.

Weakness 1. Data quality and total cost

Data quality and total cost are insufficiently established.

(a) Quality for realistic netlists

Appendix B.3 samples legal layouts and rewires connectivity to favor short connections; this does not establish quality for realistic fixed netlists.

Quality on real netlists is evaluated directly. We assess the resulting placements on the original benchmark netlists, so the evaluation makes no assumption about the quality of the synthetic training layouts. Relative to the matched no-SPA pipeline, synthetic-target SPA reduces downstream wHPWL on 22 of 25 IBM and ISPD circuits (Table 1) and improves post-route WNS and TNS on seven of eight ChiPBench circuits each (Tables 14–15).

Targets specify a spatial prior, not reference solutions. The synthetic targets are not intended as high-quality solutions for any real netlist. They specify the conditional layout distribution that SPA learns (Appendix A.1). The same formulation also accepts optimizer placements of real netlists, such as DREAMPlace results (Table 13). We have added a clarifying sentence to Section 3.1 (L166–167).

The layout-first construction follows established practice. ChipDiffusion [1] also samples a layout first and then constructs connectivity that favors it, and its models transfer zero-shot to real circuits. In FlowPlace [2], replacing random samples with boundary-aware ones improves the average downstream ranks on ICCAD 2015, from 1.75 to 1.25 for rWL and WNS and from 1.88 to 1.12 for TNS. Our Peripheral profile follows placement practice, where large macros are placed near the die periphery to preserve central space for standard cells and routing [3].

(b) Quality criteria and sensitivity

Quality criteria and sensitivity to poorer pretraining/target samples are missing.

We use three criteria, at three levels.

  • Each sample: geometric validity. Algorithm B.1 places every macro at an in-die origin that avoids the footprints already placed, so each synthetic target is legal by construction (Appendix B.3).
  • The corpus: fidelity to the prescribed profile. A spatial prior is a property of the layout distribution, so we check it on corpus averages of the bounding-box area and pairwise distance of the largest-quartile macros (Appendix D.2.2 and Table 29).
  • The adapted model: downstream utility. Whether a profile suits the objective is judged on real netlists, by downstream wHPWL (Tables 1, 12, and 13) and post-route metrics (Tables 14–15).

The first two can be checked before training. The third is known only after the downstream flow, and Table 12 shows that it is needed.

Pretraining quantity. SPA's gain holds from 5K to 40K layouts. Reducing the corpus from 40K to 5K layouts changes average Base wirelength by 2.3%, and SPA improves all four metrics at every scale. For each circuit and model, we first averaged each metric over three evaluation runs and then averaged the per-circuit values across the eight circuits. Each entry gives the Base average over the eight ChiPBench circuits, followed by the relative improvement of that average with SPA. Positive values are better, and WNS and TNS gains use violation magnitudes (Appendix F.2).

Pretraining layoutsWirelength (10⁶ μm)Congestion (%)WNS (ns)TNS (ns)
5K6.144 / +9.23%29.14 / +6.27%−1.227 / +43.71%−3999.0 / +58.23%
10K6.109 / +6.05%29.22 / +3.93%−1.086 / +30.59%−3917.9 / +35.87%
20K6.015 / +6.54%29.02 / +5.28%−1.102 / +22.63%−2268.0 / +1.78%
40K (paper)6.007 / +7.02%28.65 / +4.06%−1.066 / +26.73%−3101.4 / +15.87%

Pretraining quality. A biased corpus mainly degrades timing. We pretrained a Base on a corpus mixed with 10% strongly clustered layouts. Its downstream wHPWL stays within about 1% of the original Base on IBM and ISPD, and its geometry score on ISPD drops by 1.15 points. On ChiPBench, one circuit fails the flow, and the Base is worse in wirelength, WNS, and TNS on six of the seven circuits that complete it, with median changes of +0.8%, +15.6%, and +43.7% (Appendix F.3).

AuthorsOpenReview ↗

Response to Reviewer JJLJ – Part 2 / 3

Target samples. The spatial profile matters, and a modest corpus suffices.

  • Profile. Random and Clustered targets are our examples of poorer target samples. They share Peripheral's circuit sampling, macro sizes, degree sequences, and connectivity-generation rule. Both reach comparable or slightly lower macro-only HPWL, yet their downstream wHPWL is 1.42% to 4.14% higher (Table 12). Relative to the no-SPA pipeline, wHPWL changes by −2.20% with Random, −2.48% with Clustered, and −3.98% with Peripheral on IBM, and by +1.84%, +1.39%, and −2.19% on ISPD. A poorly chosen profile therefore reduces the gain on IBM and removes it on ISPD.
  • Coordinates under fixed netlists. On ChiPBench, we train three SPA states on the same synthetic netlists constructed from the Dispersed layouts, varying only the target macro coordinates (Section 4.4, L454–457). Compared with Compact, Relaxed reduces WNS violation magnitude on six of eight circuits and TNS violation magnitude on five. Across all eight circuits, the median reductions are 18.4% and 22.5%, respectively, with small median increases in routed wirelength and congestion. Dispersed yields the lowest routed wirelength and congestion on all eight circuits (Table 30). Relative to Compact and Relaxed, the median reductions range from 7.3% to 7.6% for routed wirelength and from 12.3% to 17.4% for congestion.
  • Source. DREAMPlace-generated targets reduce downstream wHPWL further than synthetic targets, by 3.71% on ibm01–09 and by 2.87% on ibm10–18, which are excluded from adapter training (Table 13).
  • Sample count. Among 10, 20, 50, and 100 targets per group, 20 gives the lowest downstream wHPWL in all three groups. Macro-only HPWL changes by at most about 6% relative to the 20-sample corpus, with no consistent trend. Downstream wHPWL varies more, most on Bigblue (Appendix F.4).

(c) Dataset-generation cost

Dataset-generation costs, including synthetic data and DREAMPlace teacher runs, are unreported.

Target generation takes seconds for synthetic data and minutes for DREAMPlace teachers. The table lists the full cost, which we have added to Appendix F.5. Target generation was timed on ChiPBench. Target generation and SPA training are performed once per target setting and shared by all circuits in the group (Appendix D.1.2). All reported times are averages over three runs.

StageCostFrequency
Pretraining corpus (40K layouts)41.5 minOnce
Base pretraining (3M steps)36.8 GPU-hoursOnce, shared by all benchmarks and priors
Synthetic targets (20 samples)1.03 s + 0.45 s validationOnce per target setting
DP-teacher targets (30 DREAMPlace runs)320.21 sOnce per target setting
SPA training47.5 s per 1K updates (2K–4K updates)Once per target setting

(d) Runtime and practical trade-off

Runtime comparisons with DREAMPlace and other solver-based methods are absent, leaving the practical trade-off unclear.

DREAMPlace is faster, and the gap narrows with circuit size. We have added DREAMPlace to Figure 5 and to the baseline list in Section 4.3 (L372–373). The figure already included the search-based WireMask-EA.

CircuitDREAMPlace (s)MacroSPA, sparse (s)
adaptec118.35100.07
adaptec443.11100.91
bigblue490.75201.66
bigblue2148.50270.71

Both columns report macro placement time on the same instances as Figure 5. MacroSPA times cover guided sampling, legalization, and MGR (Table 9). The MacroSPA-to-DREAMPlace runtime ratio falls from 5.5× to 1.8× across the four instances. Among learning-based placers, sparse evaluation makes MacroSPA 43.35× faster than ChipDiffusion on bigblue2.

MacroSPA’s benefit lies in an adjustable prior and downstream quality. We would like to emphasize two aspects of MacroSPA. First, the spatial prior becomes an adjustable input. After a one-time pretraining, the same frozen Base acquires a new prior from seconds of data generation and a few minutes of training for 18,562 parameters, whereas an analytical placer optimizes one fixed objective per circuit. Second, this adjustability translates into physical-design quality. On ChiPBench, DREAMPlace ranks first in routed wirelength, while synthetic SPA achieves the best overall average rank (2.90 versus 3.19 for DREAMPlace, Table 2).

AuthorsOpenReview ↗

Response to Reviewer JJLJ – Part 3 / 3

Weakness 2. Distinct contribution and parameter-matched ablations

SPA combines standard residual adapters, LoRA, and head tuning, so its distinct contribution needs stronger evidence. Downstream component ablations and comparisons with parameter-matched adaptation alternatives are missing.

Under a matched parameter budget, no alternative reproduces Full SPA's downstream gain. We agree that the adaptation components are standard. What we consider most distinctive is the formulation, in which a single pretrained placer acquires distinct spatial priors specified directly by context–placement samples, without quality scores or rankings. Within SPA, the design choice we test here is that the residual adapters take the occupancy features of the current layout as input, before graph message passing. Table 7 ablates components on macro-only metrics without controlling the trainable-parameter budget, so we added a budget-controlled downstream ablation on ChiPBench.

Setup. Three alternatives match SPA's trainable budget to within 1.1%. A and B serve as the downstream component ablations. A removes the adapters, and B removes the fusion LoRA. All variants use the same training data, initialization, training budget, sampling noise, and OpenROAD flow as SPA, and all fine-tune the output head.

VariantFusion LoRATrainable paramsTests whether SPA's gain can be obtained by
Full SPArank 818,562n/a
A: Fusion LoRA onlyrank 5418,402more LoRA capacity
B: Adapters onlynone18,766adapters alone, without the fusion LoRA
C: Generic adapters + LoRArank 818,438the same design without the occupancy input

Results.

VariantWirelengthCongestionWNSTNSNVPPowerAreaImproved pairs
Full SPA−2.9% (5)−0.2% (4)−20.2% (7)−35.0% (7)−14.7% (7)−0.5% (5)−0.4% (6)41/56
A: Fusion LoRA only−0.4% (5)+1.8% (1)−1.6% (4)−4.8% (5)+9.8% (3)+0.8% (3)+0.5% (2)23/56
B: Adapters only+0.6% (3)+0.9% (2)+6.4% (3)−14.4% (5)−2.0% (5)+0.6% (2)+0.4% (2)22/56
C: Generic adapters + LoRA+0.9% (3)−0.3% (5)−0.1% (4)−9.1% (5)0.0% (4)−0.1% (4)+0.2% (2)27/56

We evaluated each setting three times per circuit and averaged the results. Median circuit-wise change relative to the unadapted pipeline (MacroSPA w/o SPA) on eight ChiPBench circuits. Negative is better. Parentheses give the number of circuits improved. WNS and TNS are measured as violation magnitudes.

  1. Full SPA improves the most circuit–metric pairs. It improves 41 of 56 pairs over the unadapted pipeline, against 22 to 27 for the alternatives.
  2. The gap is clearest on timing. Full SPA improves WNS, TNS, and NVP on seven of eight circuits each. No alternative exceeds five of eight on any of them.
  3. The occupancy input matters at an equal budget. Variant C keeps the same two adapter locations, rank-8 fusion LoRA, and trainable output head, but feeds the node hidden state into the adapters, with the bottleneck widened from 32 to 34. Head-to-head, Full SPA is better than A, B, and C on 41, 38, and 36 of the 56 pairs. Against C, it has lower TNS violations on six of eight circuits and fewer violating paths on six of eight.

Per-circuit results are in Appendix F.1, and Section 4.2 (L323) now summarizes this ablation.

We hope these responses address your concerns. If so, we would be grateful if you would consider updating your assessment. If anything remains unclear, we are happy to continue the discussion.

[1] Chip Placement with Diffusion Models. ICML 2025.

[2] FlowPlace: Flow Matching for Chip Placement. DAC 2026.

[3] IncreMacro: Incremental Macro Placement Refinement. ISPD 2024.

Reviewer_QVoiOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MacroSPA addresses a mismatch between the spatial preferences inherited by a pretrained diffusion-based macro placer and those appropriate for a target design setting. It learns spatial preferences from synthetic or optimizer-generated context–placement pairs through spatially conditioned residual adapters, fusion LoRA, and output-head updates, while keeping the backbone frozen and training approximately 3% of the adapted model's parameters. The framework also introduces Minkowski Guided Refinement for residual geometric conflicts and sparse constraint evaluation for scalable inference. Experiments demonstrate acquisition of different spatial preferences, improvements over a matched pipeline without SPA, and competitive downstream placement and post-route results. An informative finding is that a spatial preference improving macro-only wirelength can worsen downstream mixed-size placement.

Critical strengths

1.The central claim of spatial-prior acquisition is supported by a well-targeted control. Figure 3 and Appendix D.2.2 hold circuit contexts fixed across target profiles and change only the target coordinates. Guidance and legalization are disabled during sampling, and the resulting spatial differences persist after MGR. This directly supports the interpretation that SPA changes the model's learned spatial behavior, rather than merely exploiting different connectivity inputs or postprocessing. Together with the reversal between macro-only and downstream preferences in Table 12, this gives the paper a clear and practically meaningful central insight. 2.The evidence connects adaptation to downstream utility while controlling important confounders. The matched comparisons with and without SPA share inference and downstream settings, and the component ablations in Table 7 exclude MGR. The paper reports improvements in downstream wHPWL on 22 of 25 IBM/ISPD circuits, with average reductions of approximately 3.8% and 2.2%, respectively. ChipBench extends the evaluation to routing and post-route metrics. Moreover, the DP-teacher experiment reports improvements on IBM circuits excluded from adapter training. These results make explanations based solely on postprocessing or memorization insufficient to account for the main findings.

Critical weaknesses

1.The methodological novelty is meaningful but relatively incremental. Spatially informed synthetic data, residual adapters, and LoRA are established ingredients. Appendix A provides a useful interpretation through the standard squared-loss decomposition, but does not establish that the proposed compact correction is sufficient or that reduced denoising risk guarantees improved physical-design outcomes. The component controls support the proposed design, although the downstream experiments do not completely separate the benefits of acquiring a spatial preference from broader adaptation to target-derived graph statistics. This limits the strength of the novelty claim and keeps my recommendation at weak acceptance. 2.The downstream and transfer conclusions require a carefully bounded interpretation. The principal synthetic adapters use statistics from the target circuit groups, and the ChipBench DP-teacher adapter is trained using placements from six of the eight evaluated circuits. These are legitimate adaptation settings, but the aggregate results should not be interpreted as uniformly demonstrating transfer to entirely unseen circuits. Similarly, the best average ChipBench rank does not imply consistent improvement across physical-design objectives: synthetic SPA improves WNS and TNS on seven of eight circuits relative to the no-SPA pipeline, but improves congestion on only four, and increases vga_lcd NVP from 16,329 to 19,953. The evidence supports useful, objective-dependent adaptation rather than broad downstream dominance. This is primarily a limitation of scope and interpretation, not a fundamental invalidation of the core result.

Questions for authors

1.For the ChipBench DP-teacher experiment, could you separately summarize the results on the six circuits used to generate adaptation targets and the two excluded circuits, using the existing results? The IBM holdout experiment is encouraging; this breakdown would clarify how strongly the post-route evidence supports transfer beyond the adaptation circuits. 2.In Table 12, target profiles share circuit sampling, macro-size sequences, and degree-sequence conditions, but synthetic connectivity is subsequently constructed using the sampled layouts. Can the authors clarify how much of the downstream difference can be attributed specifically to spatial preferences rather than changes in the generated connectivity distribution? Does the fixed-context experiment in Appendix D.2.2 provide any corresponding placement-quality evidence beyond the spatial-response metrics? This clarification would affect how strongly I interpret the proposed mechanism, rather than whether the reported pipeline improvements exist.

6 comments
AuthorsOpenReview ↗

Response to Reviewer QVoi – Part 1 / 3

Dear Reviewer QVoi,

Thank you for the careful and constructive review. We respond to each point below and have revised the paper accordingly.

Weakness 1. Novelty and separation of effects

(a) Established ingredients

The methodological novelty is meaningful but relatively incremental. Spatially informed synthetic data, residual adapters, and LoRA are established ingredients.

Our main contribution is the capability, and SPA is the means. We agree that residual adapters, LoRA, and spatially informed synthetic data are established ingredients. What is new is what they are used for. One pretrained placer acquires distinct spatial priors after pretraining, specified directly by context–placement samples, without quality scores or rankings and without retraining the Base. Prior generative placers fix the spatial prior when the pretraining data are constructed (FlowPlace [1], MacroDiff+ [2]) or adapt by full-model fine-tuning (ChipDiffusion [3]), as compared in Appendix E and Table 16. SPA is one compact way to realize this capability, with 18.6K trainable parameters per prior, and MGR with size-adaptive constraint evaluation makes its inference scalable. Figure 3 shows the same Base acquiring three different priors under matched contexts.

A parameter-matched ablation supports SPA's design. On ChiPBench, none of three parameter-matched alternatives reproduces SPA's downstream benefit. Full SPA improves 41 of 56 circuit–metric pairs over the unadapted baseline. Against the same baseline, fusion LoRA alone at a higher rank improves 23 of 56 pairs, adapters without fusion LoRA improve 22 of 56, and adapters using hidden states in place of occupancy features improve 27 of 56. Section 4.2 (L323) summarizes this ablation, and Appendix F.1 reports the full results.

(b) Role of Appendix A

Appendix A provides a useful interpretation through the standard squared-loss decomposition, but does not establish that the proposed compact correction is sufficient or that reduced denoising risk guarantees improved physical-design outcomes.

Appendix A defines what SPA should learn. This is also how we view Appendix A. Under the standard denoising loss, the target samples define the correction to the pretrained denoiser that SPA aims to learn. Lower target denoising risk relative to Base indicates a closer mean square approximation to the target residual (Table 5). Appendix D.2.4 compares SPA with full fine-tuning on placement quality and training cost. Tables 1, 2, and 12 evaluate downstream utility. We have added a sentence to Section 3.1 (L207–208) that states what the risk measures.

(c) Spatial preference vs. target statistics

The component controls support the proposed design, although the downstream experiments do not completely separate the benefits of acquiring a spatial preference from broader adaptation to target-derived graph statistics.

Target statistics alone do not produce consistent gains. Random uses target-derived statistics without imposing a spatial preference. Clustered favors the interior, while Peripheral favors placing large macros near the die boundary. For each profile, we report the change in downstream wHPWL relative to the same no-SPA baseline on IBM and ISPD.

Downstream wHPWL vs. no-SPAIBMISPD
Random SPA (no spatial preference)−2.20% (10/17)+1.84% (3/8)
Clustered SPA (interior preference)−2.48% (10/17)+1.39% (4/8)
Peripheral SPA (peripheral preference)−3.98% (14/17)−2.19% (8/8)

Average of circuit-wise changes in wHPWL, as in Table 12. Parentheses give the number of circuits improved.

Peripheral is the only profile that reduces average downstream wHPWL relative to no-SPA on both IBM and ISPD. Random and Clustered also improve IBM, but both increase wHPWL on ISPD. All three profiles use the same circuit sampling, macro-size sequences, degree sequences, and connectivity-generation rule. Because that rule uses the sampled coordinates, the resulting netlists can differ across profiles. Question 2 tests the effect of target coordinates with the netlists held fixed.

AuthorsOpenReview ↗

Response to Reviewer QVoi – Part 2 / 3

Weakness 2. Scope of the conclusions

(a) Transfer to unseen circuits

The downstream and transfer conclusions require a carefully bounded interpretation. The principal synthetic adapters use statistics from the target circuit groups, and the ChipBench DP-teacher adapter is trained using placements from six of the eight evaluated circuits. These are legitimate adaptation settings, but the aggregate results should not be interpreted as uniformly demonstrating transfer to entirely unseen circuits.

We agree with this bounded reading. Both are adaptation settings, as described in Appendices D.1.2 and D.4.4, and the aggregate results are not meant as evidence of transfer to entirely unseen circuits. Evidence beyond the adaptation set is reported separately, in the IBM holdout (Table 13) and for the two excluded ChiPBench circuits (Question 1).

(b) Objective-dependent improvements

Similarly, the best average ChipBench rank does not imply consistent improvement across physical-design objectives … The evidence supports useful, objective-dependent adaptation rather than broad downstream dominance.

We agree that SPA’s gains depend on the metric. Section 4.4 (L472–474) now notes this trade-off. As Tables 14–15 show, synthetic SPA improves WNS and TNS on seven of eight circuits each, while congestion improves on four and NVP increases on vga_lcd.

No method or target prior leads on every metric. In Table 2, DREAMPlace ranks first on wirelength and congestion and ties with ChiPFormer for last on power (4.69). The same holds between SPA states on one frozen Base: DP-teacher ranks better on WNS (2.25 vs. 2.50), whereas synthetic SPA ranks better on TNS, NVP, power, and area. Among synthetic profiles, Random and Clustered reach lower macro-only HPWL, yet Peripheral gives lower downstream wHPWL on both IBM and ISPD (Table 12).

This makes the target samples a practical handle on the trade-off. Each SPA state has only 18.6K trainable parameters, so engineers can keep several states and rank the resulting layouts under their own metric weights (see our response to Reviewer 8Vot, Weakness 2).

Question 1. Adaptation vs. excluded circuits

For the ChipBench DP-teacher experiment, could you separately summarize the results on the six circuits used to generate adaptation targets and the two excluded circuits, using the existing results? The IBM holdout experiment is encouraging; this breakdown would clarify how strongly the post-route evidence supports transfer beyond the adaptation circuits.

DP-teacher SPA improves six of seven metrics on each excluded circuit. Compared with the matched no-SPA baseline, ariane81 and bp68 both improve in wirelength, congestion, WNS, TNS, NVP, and power. Standard-cell area increases by less than 0.1% on both circuits. Across the six circuits used to generate adaptation targets, DP-teacher SPA improves 24 of 42 circuit–metric pairs (Tables 14–15). Section 4.4 (L474–477) now summarizes the results for both groups.

CircuitsWLCong.WNSTNSNVPPowerAreaTotal
Excluded (ariane81, bp68)2/22/22/22/22/22/20/212/14
Adaptation (six circuits)2/62/65/66/63/63/63/624/42

Post-route evaluation is not available for the IBM holdout. We would have liked to report post-route metrics there as well. However, IBM and ISPD2005 provide placement geometry and connectivity only, without the Liberty timing libraries, functional netlists, and SDC constraints that routing and timing analysis require [4]. The IBM holdout is therefore reported with downstream wHPWL, and post-route transfer is assessed on the excluded ChiPBench circuits.

AuthorsOpenReview ↗

Response to Reviewer QVoi – Part 3 / 3

Question 2. Spatial preference vs. generated connectivity

In Table 12, target profiles share circuit sampling, macro-size sequences, and degree-sequence conditions, but synthetic connectivity is subsequently constructed using the sampled layouts. Can the authors clarify how much of the downstream difference can be attributed specifically to spatial preferences rather than changes in generated connectivity distribution? Does the fixed-context experiment in Appendix D.2.2 provide any corresponding placement-quality evidence beyond the spatial-response metrics?

Yes. With netlists held fixed, the target coordinates still change post-route outcomes. Table 12 alone cannot quantify how much of the downstream difference comes from the spatial preference, because connectivity is generated from each sampled layout and the two change together. We therefore applied the fixed-context protocol of Appendix D.2.2 to ChiPBench. Three SPA states were trained on identical synthetic contexts for the ChiPBench circuits, differing only in the target coordinates, and each was run through the full downstream flow on the eight circuits. For each circuit and profile, we averaged each metric over three evaluation runs before computing the pairwise percentage changes. Each entry gives the number of circuits on which the first profile is better, followed by the median change. Negative values favor the first profile.

ComparisonWLCong.WNSTNS
Relaxed vs. Compact4/8, +0.4%4/8, +0.5%6/8, −18.4%5/8, −22.5%
Dispersed vs. Compact8/8, −7.6%8/8, −12.3%5/8, −17.2%5/8, −31.9%
Dispersed vs. Relaxed8/8, −7.3%8/8, −17.4%4/8, +3.3%4/8, −14.3%

Relaxed and Compact give the cleanest comparison. All three profiles share the same netlists, which were built from the Dispersed layouts. Compact and Relaxed therefore differ only in their spatial profile, whereas Dispersed also matches the netlists it is trained with. For each circuit, we computed the percentage change for Relaxed relative to Compact. Across the eight circuits, the median changes were +0.4% for routed wirelength, +0.5% for congestion, −18.4% for WNS violation magnitude, and −22.5% for TNS violation magnitude.

Dispersed yields the lowest routed wirelength and congestion. It improves both metrics over Compact and Relaxed on all eight circuits. Section 4.4 (L454–457) summarizes the experiment, and Appendix F.6 reports the per-circuit results.

We hope these responses address your concerns. If so, we would be grateful if you would consider updating your assessment. If anything remains unclear, we are happy to continue the discussion.

[1] FlowPlace: Flow Matching for Chip Placement. DAC 2026.

[2] Physics-Guided Geometric Diffusion for Macro Placement Generation. IJCAI 2026.

[3] Chip Placement with Diffusion Models. ICML 2025.

[4] Benchmarking End-to-End Performance of AI-Based Chip Placement Algorithms. NeurIPS 2025.

Reviewer_8VotOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes MacroSPA, a post-training framework to adaptively learn spatial prior from target placement samples, using pretrained diffusion-based macro placer. Specifically, it uses randomly sampled legal layouts for pretraining following ChipDiffusion, and then fine-tunes using larger synthetic circuits. It then adopts Minkowski guided refinement to reduce overlap. Results show that MacroSPA can optimize not only HPWL and RUDY, but also post-route PPA.

Critical strengths
  • This paper considers optimizing actual PPA beyond HPWL, which is a practical promblem.
  • The idae of post-training a placer using high-quality placement samples makes sense.
Critical weaknesses
  • Does pretraining matter? Can the authors design some experiments to show the scaling law of data colume?
  • Is it possible to contral the weights of different metrics? I think it may be hard. But in practice engineers will need to adopt different weights to select from different layouts.
  • In practice it may be hard to obtain many high-quality samlpes for fine-tuning, especially those industrial chips. So it may be very difficult to generalize to unseen chips.
  • Is it possible to compare the proposed method with commercial placement tools?
Questions for authors

See weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 8Vot – Part 1 / 3

Dear Reviewer 8Vot,

Thank you for the positive assessment and the practical questions. We first address one point in the summary and then respond to each weakness.

We have clarified the adaptation setting in Section 4.1. The original description could suggest that MacroSPA fine-tunes the model on larger synthetic circuits, and we have revised it to make the setting explicit. The pretrained backbone stays frozen during SPA training. SPA trains only two small residual adapters, a low-rank update of the fusion layer, and the output head, which together have 18.6K parameters, about 3% of the adapted model. In our primary synthetic setting, they are trained on 20 target samples per circuit group within a few minutes. Section 4.1 now states the pretraining setup (L270–272) and the 20-pair adaptation protocol (L275–276).

Weakness 1. Role of pretraining and data scaling

Does pretraining matter? Can the authors design some experiments to show the scaling law of data volume?

Yes, pretraining matters, through both the size and the composition of its corpus. We tested both on our Spatial Base. Prior work also shows that the pretraining distribution shapes learned placement behavior. ChipDiffusion reports higher zero-shot legality on clustered IBM with its broader v1 training distribution (0.8213) than with v0 (0.7794) [1], and FlowPlace reports better downstream ranks with boundary-aware synthetic layouts than with random ones [2]. As discussed in Appendix E, these methods fix the spatial prior when constructing the pretraining data, whereas MacroSPA keeps a general pretrained Base and revises its prior afterward with SPA.

Data volume. Base wirelength falls as the pretraining corpus grows. We pretrained Bases on nested random subsets of 5K, 10K, and 20K layouts from the original 40K corpus. All four Bases used the same initial weights, validation set, and budget of 3M updates. We then applied SPA with the same synthetic targets and evaluated on ChiPBench. Each cell reports the Base metric, averaged first over three evaluation runs per circuit and then across the eight circuits, followed by the relative improvement of that average with SPA. Positive gains indicate improvement. For WNS and TNS, gains are computed on violation magnitudes.

Pretraining dataWirelength (10⁶ μm)Congestion (%)WNS (ns)TNS (ns)
5K6.144 / +9.23%29.14 / +6.27%−1.227 / +43.71%−3999.0 / +58.23%
10K6.109 / +6.05%29.22 / +3.93%−1.086 / +30.59%−3917.9 / +35.87%
20K6.015 / +6.54%29.02 / +5.28%−1.102 / +22.63%−2268.0 / +1.78%
40K (paper)6.007 / +7.02%28.65 / +4.06%−1.066 / +26.73%−3101.4 / +15.87%

Average Base wirelength decreases monotonically from 6.144 to 6.007 × 10⁶ μm. WNS improves overall, while the 20K Base has the smallest TNS violation magnitude. SPA improves all four average metrics at every scale, and its gain is largest for the 5K Base, which suggests that SPA partly compensates for weaker pretraining. At each scale, synthetic SPA changes average power and standard-cell area by less than 2% relative to the corresponding Base (Appendix F.2).

Data quality. A biased corpus is inherited by the Base. We pretrained a Base on a corpus in which 10% of the layouts are strongly clustered. Its downstream wHPWL on IBM and ISPD stays within about 1% of the original Base, but its geometry score on ISPD drops by 1.15 points. On ChiPBench, it is worse in wirelength, WNS, and TNS on six of the seven circuits that complete the flow (Appendix F.3).

AuthorsOpenReview ↗

Response to Reviewer 8Vot – Part 2 / 3

Weakness 2. Controlling the weights of different metrics

Is it possible to control the weights of different metrics? I think it may be hard. But in practice engineers will need to adopt different weights to select from different layouts.

Metric weights can guide layout selection. Engineers can generate candidate layouts with different SPA states and rank them by a weighted sum of normalized metrics. The weights apply at selection time. During sampling, the choice of SPA state shapes the trade-off. For example, DP-teacher SPA ranks better on WNS, while synthetic SPA ranks better on TNS and standard-cell area (Table 2).

Switching SPA states during sampling is a preliminary route to finer control. Both states share one frozen backbone, so they can be switched within a single reverse-diffusion trajectory without retraining. Assuming that high-noise steps shape the global organization and low-noise steps refine it, we start with the synthetic state and switch to the DP-teacher state halfway through reverse diffusion on all three circuits.

CircuitWirelengthCongestionWNSTNSNVPPowerArea
ariane81−9.77%−9.69%−45.65%−25.53%−32.21%−0.28%−0.63%
ariane133−0.50%−0.45%−20.77%−24.55%−28.60%+0.26%+0.19%
vga_lcd+1.85%+1.83%−15.89%−11.41%+6.32%+0.63%+0.62%

For each circuit and metric, we used the three-run average for each setting to compute the percentage change relative to DP-teacher-only sampling. Negative values indicate improvement. WNS and TNS are computed on violation magnitudes.

Switching reduces WNS and TNS violation magnitudes on all three circuits, while its effect on wirelength and congestion depends on the circuit. On vga_lcd, for example, switching reduces the WNS violation magnitude by 15.89% at the cost of a 1.85% increase in routed wirelength relative to DP-teacher-only sampling. A natural next step is to learn or search for the schedule that best matches a given set of metric weights, once its behavior has been validated across more schedules and circuits. We include this direction in the revised conclusion (Section 5, L497–499).

AuthorsOpenReview ↗

Response to Reviewer 8Vot – Part 3 / 3

Weakness 3. Sample availability and generalization

In practice it may be hard to obtain many high-quality samples for fine-tuning, especially those industrial chips. So it may be very difficult to generalize to unseen chips.

In MacroSPA, high-quality real placements are an option, not a requirement. The Spatial Base is pretrained only on synthetic layouts and already places real IBM and ISPD circuits zero-shot with quality comparable to ChipDiffusion (Table 1, w/o SPA) [1]. SPA can then adapt it without such placements as well. As Appendix A.1 shows, the prior used in post-training can come from two sources. Optimizer-generated placements, such as DREAMPlace results, define it through the netlist-first factorization. Synthetic layouts that encode only a coarse spatial intent, such as placing large macros near the periphery, define it through the layout-first factorization. The intent need not be precise, only appropriate for the target objective (Table 12). Synthetic targets require only the macro sizes and degree statistics of the target circuit group. Because only aggregate statistics are used, this can also ease confidentiality constraints in industrial settings.

Twenty targets give the lowest downstream wHPWL in all three groups. We tested 10, 20, 50, and 100 targets, training one adapter per group at each target count. At 20 targets, macro HPWL and RUDY come from the Full SPA row of Table 7. Macro-only HPWL and RUDY show no consistent trend with sample count, while downstream wHPWL varies most across counts on Bigblue.

GroupSamplesHPWL (10⁵)RUDY (10²)wHPWL (10⁶)
IBM G1101.491.336.55
201.431.276.10
501.451.316.82
1001.431.306.93
Adaptec1034.514.70126.56
2036.364.73123.59
5036.734.58125.85
10034.154.68125.29
Bigblue1050.115.75431.32
2051.515.76355.99
5052.406.15386.73
10049.165.64394.89

DP-teacher SPA also improves results on circuits excluded from adaptation. Relative to the no-SPA baseline, it reduces average downstream wHPWL by 6.79% on the held-out IBM circuits (Table 13) and improves six of seven metrics on each of the two excluded ChiPBench circuits (Section 4.4, L475–477, and Tables 14–15). For a new chip, synthetic targets can likewise be constructed from its netlist statistics without an optimized reference layout, which provides a practical route to target-aware adaptation when high-quality placements are unavailable.

Weakness 4. Comparison with commercial tools

Is it possible to compare the proposed method with commercial placement tools?

We currently do not have access to commercial placement tools, so a direct comparison is not possible in this work. We therefore use the open-source ChiPBench setting, which provides a common OpenROAD flow for post-route comparison. Within this setting, our evaluation covers the GPU-accelerated analytical placer DREAMPlace and methods based on reinforcement learning (ChiPFormer), black-box optimization (WireMask-EA), and generative modeling (ChipDiffusion, FlowPlace, MacroDiff+).

We hope these responses address your questions. If so, we would be grateful if you would consider updating your assessment. If anything remains unclear, we are happy to continue the discussion.

[1] Chip Placement with Diffusion Models. ICML 2025.

[2] FlowPlace: Flow Matching for Chip Placement. DAC 2026.

AuthorsOpenReview ↗

We look forward to your feedback

Dear Reviewer 8Vot,

Thank you again for your time and effort in reviewing our paper. As the discussion period closes on November 18, we would like to check whether our response has addressed your concerns. In brief, we added pretraining experiments on corpus size and composition (Appendices F.2–F.3), a target-count sensitivity analysis (Appendix F.4), and preliminary results on switching SPA states to trade off metrics.

If our response has addressed your questions, we would be grateful if you would consider updating your assessment. If not, please let us know what remains unclear, and we will respond promptly.

Best regards, Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

Dear Reviewers and Area Chair,

Thank you for your careful reviews. We highlight main text changes in red and report the new experiments in Appendix F.

Parameter-matched SPA ablation (JJLJ, QVoi). Against the same no-SPA baseline, Full SPA improves 41 of 56 circuit–metric pairs on eight ChiPBench circuits. Three alternatives matched within 1.1% of its trainable parameter count improve 22–27 pairs. Full SPA improves WNS, TNS, and the number of violating paths on seven of eight circuits each. Compared with the closest alternative, whose adapters use node hidden states instead of occupancy features, Full SPA has smaller TNS violations on six of eight circuits (Section 4.2, Appendix F.1).

Effect of target coordinates (QVoi, JJLJ). We extended Figure 3's fixed-context experiment to post-route evaluation on eight ChiPBench circuits. Only the training coordinates change, the synthetic netlists are held fixed, and guidance and legalization are disabled during macro placement. Relative to Compact, Relaxed reduces WNS and TNS violation magnitudes by medians of 18.4% and 22.5%, respectively, with small median increases in routed wirelength and congestion.

Transfer to circuits excluded from adaptation (QVoi). On each ChiPBench circuit excluded from DP-teacher adaptation (ariane81 and bp68), DP-teacher SPA improves six of seven metrics over the no-SPA baseline, and standard-cell area increases by less than 0.1% (Section 4.4, Tables 14–15).

Training data (8Vot, JJLJ). SPA improves average wirelength, congestion, WNS, and TNS on ChiPBench at all four pretraining corpus sizes tested (5K–40K layouts). Mixing 10% strongly clustered layouts into pretraining worsens the unadapted Base's WNS and TNS on six of the seven circuits that complete the flow. Among the four target counts tested (10, 20, 50, 100), 20 gives the lowest downstream wHPWL in all three groups (Appendices F.2–F.4).

Cost and runtime (JJLJ). On ChiPBench, generating 20 synthetic targets takes 1.5 s including validation, and generating 30 DREAMPlace teacher targets takes 320 s. After one-time Base pretraining, an SPA state trains in 1.6–3.2 minutes (Appendix F.5). Figure 5 now includes DREAMPlace, which is faster than sparse MacroSPA on all four timed ISPD instances, with the gap narrowing from 5.5× to 1.8× as circuit size grows.

Other revisions. Section 3.1 clarifies how synthetic targets specify spatial preferences and what target denoising risk measures. Section 4.1 states that the main MacroSPA variants share one pretrained Base and omit ChipDiffusion's subsequent full-model fine-tuning. Section 4.4 discusses trade-offs across post-route metrics. The conclusion describes switching SPA states during sampling as a possible way to account for designer specified metric weights in future work.

We address each reviewer's questions in detail in the individual replies.

Dear Reviewers,

The author-reviewer discussion period will end soon. Please review and acknowledge the authors’ responses.

Your AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.