acceptodds
Under review as a conference paper at ICLR 2027

EGGPlacer: Energy-Guided Generative Chip Placement with Diffusion Models

Abstract

Chip placement strongly affects routed wirelength and thermal behavior, yet these downstream metrics are difficult to evaluate during macro placement. Learning-based macro placers commonly optimize the macro-only half-perimeter wirelength (HPWL), denoted mHPWL, which ignores standard cell connectivity and spatial demand. In contrast, analytical mixed-size placers optimize differentiable full netlist wirelength and density objectives but are typically unaware of thermal effects. We propose EGGPlacer, an energy-guided diffusion framework that incorporates analytical placement feedback and thermal feedback into macro placement generation. An analytical placer acts as an online physical response model, exposing the standard cell distribution and a full netlist wirelength proxy induced by each intermediate macro placement, while a differentiable thermal model provides temperature feedback. These signals are formulated as composable energies and injected throughout reverse diffusion without retraining the underlying generative model. Experiments on public placement benchmarks show that EGGPlacer achieves competitive mixed-size HPWL relative to strong analytical and learning-based macro placers. In the end-to-end evaluation, EGGPlacer also obtains the lowest average routed wirelength and peak temperature among the evaluated methods. Ablation studies further demonstrate that analytical and thermal guidance provide distinct and complementary benefits across wirelength proxies, routed wirelength, and thermal objectives.

Primary area:
generative models
Keywords:
macro placement, diffusion models, chip design, electronic design automation, multi-objective optimization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_RSpkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EGGPlacer adds guidance to a diffusion macro placer at inference time, without retraining. At each DDIM step, the predicted clean placement is refined for a few steps on a composite energy. The energy has three parts: (i) a full-netlist WA wirelength computed after running a few DREAMPlace Nesterov steps for the standard cells with macros fixed, (ii) a smooth overlap term between macros and obstacle rectangles from dense standard-cell regions, and (iii) a thermal term from a learned set-based surrogate of HotSpot. The correction is added back into the DDIM update, and a greedy legalizer runs at the end. On 20 ChiPBench designs with the ORFS flow and HotSpot, EGGPlacer has the lowest mean ORFS HPWL, routed wirelength (2.18% below RollPlace) and peak temperature (4.90% lower temperature rise than ChipDiffusion). DREAMPlace remains better on HPWL after DREAMPlace refinement. The paper also shows that HPWL from ORFS tracks routed wirelength better than HPWL from DREAMPlace (Spearman 0.918 vs 0.660).

Critical strengths

Strength 1: The evaluation goes past macro-only HPWL. All methods are run through the same ORFS placement and routing flow and through HotSpot on 20 public designs. The observation that mHPWL, DREAMPlace HPWL and routed wirelength can rank methods differently (Tables 1, 3 and 4) is useful for the learning-based placement community, which mostly reports mHPWL.

Strength 2: The guidance mechanism is simple and practical. It works on a pretrained diffusion model without retraining, it treats the analytical placer as a black-box response with stop-gradients, and the total runtime (about 95 s per design) is much lower than EfficientPlace, ChipDiffusion and RollPlace (Table 2).

Critical weaknesses

Weakness 1: The thermal part is weakly supported. The surrogate is trained on 1,200 HotSpot placements drawn from the same 20 benchmarks used for evaluation, partly generated by earlier EGGPlacer versions, and the split is by placement, not by design (Appendix F.1). So generalization to unseen designs is not tested. The measured effect is also very small. Under joint guidance, HotSpot shows a 0.278 °C peak reduction, while the surrogate predicts 1.020 °C (Appendix F.5). Relative to the unguided variant, the full method lowers the temperature rise by only 2.0%. In addition, macro power is scaled by 10x to create hotspots, and none of the baselines optimize temperature, so the thermal comparison is one-sided.

Weakness 2: The wirelength gains are small and not tested for significance. Routed wirelength is 4.99 vs 5.10 for RollPlace, and the mean rank is 2.80 vs 2.85 for FlowPlace (Table 3). EGGPlacer is averaged over 3 seeds, but every baseline is a single run, and no per-design variance or test is given. Means over 20 designs of very different sizes are dominated by the largest designs. EGGPlacer also calls DREAMPlace inside every reverse step, yet its HPWL after DREAMPlace refinement is still worse than plain DREAMPlace (6.08 vs 5.64). Timing, DRC violations and congestion from ORFS are not reported, even though the end-to-end flow produces them.

Questions for authors
  1. How does the thermal surrogate perform with a design-level split, that is, trained on some designs and tested on unseen ones? Do the thermal results in Table 3 hold when the surrogate has never seen the evaluated design?
  2. Can you report per-design standard deviations for EGGPlacer, multiple runs for the stochastic baselines, and a paired test for routed wirelength and peak temperature?
  3. What are WNS, TNS, DRC count and routing overflow from ORFS for each method? A 2% routed wirelength gain is hard to interpret without these.
  4. Were the guidance weights and schedules tuned on the same 20 designs used for evaluation?
  5. Would a simple thermal-aware baseline, for example DREAMPlace with a macro power density term, give a similar temperature reduction?

Suggestions for future work:

Evaluating with the original macro power (without the 10x scaling) and on designs with real high-power macros would make the thermal claim more relevant to practice.

4 comments
AuthorsOpenReview ↗

Response 1 to Reviewer RSpk

Dear Reviewer RSpk,

Thank you for your detailed feedback on thermal generalization, statistical reliability, and downstream placement quality. We will upload a revised manuscript with all revisions highlighted in blue for ease of reference.

W1

Weakness 1: The thermal part is weakly supported. The surrogate is trained on 1,200 HotSpot placements drawn from the same 20 benchmarks used for evaluation, partly generated by earlier EGGPlacer versions, and the split is by placement, not by design (Appendix F.1). So generalization to unseen designs is not tested. The measured effect is also very small. Under joint guidance, HotSpot shows a 0.278 °C peak reduction, while the surrogate predicts 1.020 °C (Appendix F.5). Relative to the unguided variant, the full method lowers the temperature rise by only 2.0%. In addition, macro power is scaled by 10x to create hotspots, and none of the baselines optimize temperature, so the thermal comparison is one-sided.

Response

We thank the reviewer for highlighting the need for stronger evidence on thermal generalization, surrogate calibration, and thermal-aware comparisons. For the accuracy, evaluation, and experimental setup of the thermal model, please refer to our responses to W1 and W2 of Reviewer AsWA. For the marginal thermal benefit and its uncertainty, please see Q2 of Reviewer wT1b. For the thermal-aware refinement comparison, please see Q1 of Reviewer wT1b.

W2

Weakness 2: The wirelength gains are small and not tested for significance. Routed wirelength is 4.99 vs 5.10 for RollPlace, and the mean rank is 2.80 vs 2.85 for FlowPlace (Table 3). EGGPlacer is averaged over 3 seeds, but every baseline is a single run, and no per-design variance or test is given. Means over 20 designs of very different sizes are dominated by the largest designs. EGGPlacer also calls DREAMPlace inside every reverse step, yet its HPWL after DREAMPlace refinement is still worse than plain DREAMPlace (6.08 vs 5.64). Timing, DRC violations and congestion from ORFS are not reported, even though the end-to-end flow produces them.

Response

We appreciate the reviewer’s concerns about the reliability and practical significance of the wirelength improvements. We address the raised concerns point by point below.

First, while the routed wirelength improvements are modest, our method achieves substantially better runtime and temperature compared with RollPlace, and a clear thermal advantage over FlowPlace, while also achieving a lower mean routed wirelength. Moreover, FlowPlace relies on handcrafted priors that push macros toward the chip periphery, which tends to induce hotspots.

Second, we will add results from three seeds for every baseline in the revised manuscript.

Third, to mitigate the influence of design size on aggregate comparisons, Table 3 reports mean ranks across designs. For additional statistical comparisons, please refer to our responses to Q2 and Q3.

Fourth, EGGPlacer invokes DREAMPlace to supply analytical feedback during diffusion sampling. Our goal is not to beat standalone DREAMPlace on HPWL. Instead, we aim to combine multiple physical objectives via composable energy guidance. We emphasize that placement quality should be judged by routed wirelength from the unified downstream OpenROAD flow, rather than by HPWL alone as an intermediate proxy metric.

Finally, we will supplement the missing metrics including timing, DRC violations, and congestion from the ORFS flow in the revised manuscript.

AuthorsOpenReview ↗

Response 2 to Reviewer RSpk

Q1

How does the thermal surrogate perform with a design-level split, that is, trained on some designs and tested on unseen ones? Do the thermal results in Table 3 hold when the surrogate has never seen the evaluated design?

Response

We thank the reviewer for asking whether thermal prediction and guidance remain effective on designs excluded from surrogate training. For the accuracy, evaluation, and experimental setup of the thermal model, please refer to our responses to W1 and W2 from Reviewer AsWA. Our design-held-out experiments show mixed results: thermal guidance improves peak temperature on bp_multi57 but slightly degrades it on VeriGPU. Therefore, consistent thermal improvements on unseen designs cannot yet be established.

Q2

Can you report per-design standard deviations for EGGPlacer, multiple runs for the stochastic baselines, and a paired test for routed wirelength and peak temperature?

Response

We thank the reviewer for requesting multiple-run results and paired statistical comparisons to assess the reliability of the reported improvements.

First, per-design standard deviations for EGGPlacer are provided in Tables 7–10 of our original manuscript (Table 10-13 in revised manuscript).

Second, we will add multi-seed evaluations for all stochastic baselines. The updated mean ranks for EGGPlacer, FlowPlace, and RollPlace are shown below. For each design, we average each method’s three results, rank all six methods, and then average the ranks over the 20 designs. Lower ranks are better. The full multi-run results will be updated in the revised manuscript.

MethodRouted WL Mean Rank ↓Peak T Mean Rank ↓
EGGPlacer2.6501.600
FlowPlace2.8504.150
RollPlace3.3503.150

Third, regarding the paired statistical test, we conducted systematic paired comparisons for the full EGGPlacer pipeline, the thermal-only variant (woa, T-only), and the analytical-only variant (wot, A-only) against the FlowPlace and RollPlace baselines, covering both routed wirelength and peak temperature. For the statistical analysis, we first average the results from three random seeds for each method on each design. We then compute the mean difference across all 20 benchmark designs, along with 95% confidence intervals (CIs) derived from the Student’s t-distribution for paired samples. In the table below, each entry presents the mean reduction [95% CI], where the difference is defined as baseline method − evaluated method.

Baseline MethodEvaluated MethodRouted WL Reduction (10⁶ μm)Peak T Reduction (°C)
FlowPlaceEGGPlacer (A+T)0.3147 [−0.3097, 0.9392]2.377 [0.975, 3.779]
FlowPlacewoa (T-only)−0.1013 [−0.2312, 0.0286]2.682 [1.415, 3.949]
FlowPlacewot (A-only)0.3635 [−0.2859, 1.0128]2.099 [0.848, 3.350]
RollPlaceEGGPlacer (A+T)0.2596 [0.0534, 0.4658]1.326 [0.560, 2.093]
RollPlacewoa (T-only)−0.1565 [−0.7155, 0.4026]1.631 [0.856, 2.406]
RollPlacewot (A-only)0.3083 [0.0847, 0.5320]1.048 [0.443, 1.653]

For the targeted paired comparison between analytical-only guidance and joint analytical–thermal guidance, please refer to our response to Q2 from Reviewer wT1b.

AuthorsOpenReview ↗

Response 3 to Reviewer RSpk

Q3

What are WNS, TNS, DRC count and routing overflow from ORFS for each method? A 2% routed wirelength gain is hard to interpret without these.

Response

We appreciate the request to evaluate timing and routability alongside routed wirelength. We did not incorporate dedicated energy terms targeting WNS, TNS, DRC count, or routing overflow in our formulation, and these metrics were therefore not reported in the main manuscript. Nevertheless, EGGPlacer remains highly competitive on these ORFS timing and routability metrics. We will add these results in the appendix, including the three-seed mean values and mean ranks over the 20 benchmarks.

For each metric, we average results across three seeds per benchmark. We rank the methods within each benchmark, average tied ranks, and then aggregate the values and ranks across the 20 benchmarks. Bold font denotes the best mean value and the best mean rank for each metric.

ORFS timing and routability results across 20 benchmarks. Each cell shows the mean value followed by the mean rank in parentheses.

MethodWNS ↑ (ns)TNS ↑ (ns)DRC count ↓ (×10³)Routing overflow ↓ (×10³)
DREAMPlace−1.526 (3.250)−1246.37 (3.200)60.794 (5.200)63.724 (3.775)
EfficientPlace−1.803 (4.200)−1816.38 (4.400)1377.387 (4.000)252.625 (3.975)
ChipDiffusion−1.658 (4.050)−1709.23 (4.450)464.325 (2.700)160.829 (3.800)
RollPlace−1.684 (3.800)−1281.32 (3.500)155.848 (4.150)61.402 (3.650)
FlowPlace−1.648 (2.750)−1742.75 (2.800)253.491 (2.650)90.552 (2.800)
EGGPlacer−1.479 (2.950)−1123.13 (2.650)144.662 (2.300)49.830 (3.000)

Q4

Were the guidance weights and schedules tuned on the same 20 designs used for evaluation?

Response

We thank the reviewer for requesting clarification of how the guidance weights and schedules were selected. We used the same 20 ChiPBench designs as in the evaluation to guide our selection of guidance strength, as we did not have a separate public benchmark suitable for validation in our evaluation flow. Given the computational cost of downstream ORFS evaluation, we performed only coarse tuning of the guidance strength. More extensive tuning might yield further improvements. All four EGGPlacer variants share the same frozen pretrained diffusion backbone and use the same hyperparameter settings across the three seeds and all designs.

Q5

Would a simple thermal-aware baseline, for example DREAMPlace with a macro power density term, give a similar temperature reduction?

Response

We thank the reviewer for suggesting a simpler thermal-aware baseline. Rather than adding a macro power density penalty directly to DREAMPlace, we evaluate a thermal-aware refinement baseline initialized from DREAMPlace placements using the same differentiable thermal surrogate. Please refer to our response to Q1 of Reviewer wT1b for the experimental setup, results, and paired comparisons with woa variant and the complete EGGPlacer. These results have also been incorporated into Appendix E.4 of the revised manuscript.

AuthorsOpenReview ↗

Kind Follow-up on Our Rebuttal Submission

Dear Reviewer RSpk,

We hope you are doing well. We wanted to kindly let you know that we have submitted our rebuttal addressing your questions and concerns in detail, including:

  • Additional design-held-out thermal surrogate experiments, examining cross-design generalization and its limitations.
  • Expanded multi-seed evaluations and paired statistical comparisons, including routed wirelength and peak temperature results with 95% confidence intervals.
  • Additional downstream timing and routability evaluations, reporting WNS, TNS, DRC violations, and routing overflow across 20 benchmarks.
  • A new DREAMPlace-based thermal-aware refinement baseline, together with clarifications on guidance hyperparameter selection and experimental settings.

If you have a moment, we would be very grateful if you could review our responses and let us know whether any concerns remain that we could further clarify.

We sincerely appreciate the time and effort you have invested in reviewing our work, and thank you for considering our additional evidence and clarifications in your evaluation.

Best regards,

The Authors

Reviewer_wT1bOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EGGPlacer guides diffusion-based macro placement using DREAMPlace feedback and a learned thermal surrogate. Experiments on 20 ChiPBench designs evaluate mixed-size HPWL, routed wirelength, and HotSpot temperatures. Although DREAMPlace achieves better HPWL_DP, EGGPlacer reports better average routed wirelength and peak temperature under the common downstream flow.

Critical strengths
  • The downstream evaluation is a strength: Table 3 reports average routed wirelength of 4.99 versus 5.48 for DREAMPlace and 5.10 for RollPlace, in units of 10^6 μm. These results go beyond macro-only proxy improvements.

  • The A/T ablations use the same frozen diffusion checkpoint, and final temperatures are independently evaluated with HotSpot. The analysis of proxy rankings and surrogate miscalibration provides useful diagnostics.

Critical weaknesses
  • An appropriate comaprison with thermal-aware placers is missing. The external baselines lack explicit thermal objectives, so outperforming them in temperature does not establish competitiveness with thermal-aware optimization. ThePlace (ASP-DAC 2025) is relevant analytical work, although its global-placement scope requires adaptation or discussion. More directly, applying the same energies and thermal surrogate to an existing placement at matched runtime would test the value of guidance throughout diffusion. The current A/T ablations do not distinguish this contribution from simpler thermal-aware refinement.

  • The marginal benefit of joint guidance needs stronger validation. In Table 5, adding thermal guidance to the analytical-only variant reduces average peak temperature from 61.77°C to 61.50°C while increasing routed wirelength from 4.94 to 4.99. Appendix F.5 reports a predicted temperature reduction of 1.020°C versus 0.278°C measured by HotSpot. The measured improvement deserves credit, but the small marginal gain and surrogate overestimation make paired uncertainty estimates important for establishing a reliable wirelength–temperature tradeoff.

Questions for authors
  1. How does EGGPlacer compare with thermal-aware refinement using the same energies and surrogate, initialized from DREAMPlace or an unguided diffusion sample, at matched runtime and comparable routed-wirelength targets?

  2. Can you report paired temperature and routed-wirelength differences, with uncertainty across seeds, between analytical-only and joint guidance to establish the consistency of the marginal thermal benefit?

3 comments
AuthorsOpenReview ↗

Response 1 to Reviewer wT1b

Dear Reviewer wT1b,

Thank you for your constructive suggestions on thermal-aware comparisons and quantifying the wirelength–temperature tradeoff. We will upload a revised manuscript with all revisions highlighted in blue for ease of reference.

W1

An appropriate comaprison with thermal-aware placers is missing. The external baselines lack explicit thermal objectives, so outperforming them in temperature does not establish competitiveness with thermal-aware optimization. ThePlace (ASP-DAC 2025) is relevant analytical work, although its global-placement scope requires adaptation or discussion. More directly, applying the same energies and thermal surrogate to an existing placement at matched runtime would test the value of guidance throughout diffusion. The current A/T ablations do not distinguish this contribution from simpler thermal-aware refinement.

Response

We thank the reviewer for highlighting the need to distinguish guidance during generation from refinement afterward. We have therefore added a thermal-aware refinement baseline initialized from DREAMPlace, and its results and paired comparisons are reported in our response to Q1. These results have also been incorporated into Appendix E.4 of the revised manuscript.

We apologize that the methodological contribution of our work was not sufficiently articulated in the original submission. The core novelty of this work is the unified integration of analytical feedback and differentiable thermal feedback into the diffusion sampling pipeline, which enables downstream physical objectives to shape placement evolution across every denoising step. The thermal surrogate provides differentiable optimization signals, and our ablation studies are designed to validate the effectiveness of this stepwise guidance paradigm. This way of incorporating physical feedback also provides a basis for extending the framework to other downstream objectives. As discussed in our Conclusion, this framework holds the potential to be extended to other downstream metrics, such as total negative slack (TNS) and worst negative slack (WNS), by developing and validating suitable differentiable proxies and incorporating them as additional energy terms.

W2

The marginal benefit of joint guidance needs stronger validation. In Table 5, adding thermal guidance to the analytical-only variant reduces average peak temperature from 61.77°C to 61.50°C while increasing routed wirelength from 4.94 to 4.99. Appendix F.5 reports a predicted temperature reduction of 1.020°C versus 0.278°C measured by HotSpot. The measured improvement deserves credit, but the small marginal gain and surrogate overestimation make paired uncertainty estimates important for establishing a reliable wirelength–temperature tradeoff.

Response

We thank the reviewer for highlighting the surrogate’s overestimation and the need to quantify uncertainty in the marginal thermal benefit. We acknowledge the surrogate overestimation issue noted by the reviewer, which aligns with the experimental analysis presented in our Appendix F. To establish a more reliable characterization of the wirelength–temperature tradeoff, we have supplemented paired difference statistics along with corresponding confidence intervals for both peak temperature and routed wirelength. For detailed results, please refer to our response to Q2.

AuthorsOpenReview ↗

Response 2 to Reviewer wT1b

Q1

How does EGGPlacer compare with thermal-aware refinement using the same energies and surrogate, initialized from DREAMPlace or an unguided diffusion sample, at matched runtime and comparable routed-wirelength targets?

Response

We thank the reviewer for suggesting a comparison with thermal-aware refinement initialized from DREAMPlace.

Since routed wirelength is obtained only after downstream placement and routing, directly matching routed-wirelength targets across methods during macro placement is difficult. We therefore initialize from DREAMPlace placements and perform a fixed 200-step post-processing procedure (matching the diffusion step count of the woa variant) across all designs and seeds, using the same thermal surrogate and legality energy as the woa variant. The recorded mean total runtime is 96.32 s, comprising 19.87 s for DREAMPlace initialization and 76.44 s for refinement. We evaluate all 20 designs using three random seeds per design. The mean peak temperature and routed wirelength are 61.719°C and 5.118 × 10⁶ μm, respectively, compared with 61.190°C and 5.401 × 10⁶ μm for woa and 61.495°C and 4.985 × 10⁶ μm for the complete EGGPlacer.

For the paired comparison, we first average three seed results for each method on each design, then compute the mean difference and per-comparison 95% confidence interval (CI) across the 20 designs. We define the difference as DREAMPlace + thermal refinement − comparison method, and positive values indicate a higher value for the refinement baseline.

Comparison methodPeak temperature difference (°C), mean [95% CI]Routed WL difference (10⁶ μm), mean [95% CI]
woa+0.529 [−0.032, 1.090]−0.2830 [−0.7285, 0.1625]
Complete EGGPlacer+0.224 [−0.365, 0.814]+0.1331 [−0.1543, 0.4204]

These results show that thermal-aware refinement achieves competitive average results: relative to woa, it yields a slightly higher mean temperature and shorter mean routed wirelength, while the complete EGGPlacer achieves lower mean values for both metrics.

Q2

Can you report paired temperature and routed-wirelength differences, with uncertainty across seeds, between analytical-only and joint guidance to establish the consistency of the marginal thermal benefit?

Response

We appreciate the request for paired statistics to assess the consistency of the temperature benefit and its wirelength cost.

We conducted a paired comparison between analytical-only (A) and joint analytical–thermal guidance (A+T) across 20 ChiPBench designs and three random seeds. We define ΔT = T_peak(A) − T_peak(A+T) and ΔWL = WL(A+T) − WL(A), where positive values indicate temperature reduction and wirelength overhead, respectively. For all 95% confidence intervals (CIs) below, we first average paired differences over the three seeds within each design, then compute two-sided Student’s t intervals across the 20 designs. Mean ± SD describes variability across the three seed-level averages.

For three different seeds, the mean paired temperature reductions are 0.166, 0.316, and 0.352°C, with corresponding routed-wirelength differences of −0.0359, +0.1504, and +0.0317 × 10⁶ μm, respectively. Across seeds, joint guidance achieves a mean temperature reduction of 0.278 ± 0.099°C and a mean routed-wirelength increase of (0.0487 ± 0.0943) × 10⁶ μm. The corresponding design-level 95% CIs are [−0.136, 0.693]°C and [−0.0110, 0.1085] × 10⁶ μm, respectively. These results suggest a consistent tendency toward temperature reduction across random seeds, although the improvement varies across designs.

To further assess the contribution of thermal guidance without analytical guidance, we also compare woat and woa across the same 20 designs and three seeds. Adding thermal guidance yields a mean HotSpot peak temperature reduction of 0.702 ± 0.315°C, with a design-level 95% CI of [−0.054, 1.458]°C. The mean routed-wirelength reduction is 0.0685 × 10⁶ μm, with a corresponding design-level 95% CI of [−0.2610, 0.3980] × 10⁶ μm. Compared with the mean temperature reduction of 0.278°C from wot to EGGPlacer, this larger mean reduction suggests that thermal guidance may provide smaller additional gains when analytical guidance is already present.

AuthorsOpenReview ↗

Kind Follow-up on Our Rebuttal Submission

Dear Reviewer wT1b,

We hope you are doing well. We wanted to kindly let you know that we have submitted our rebuttal addressing your questions and concerns in detail, including:

  • A new thermal-aware refinement baseline initialized from DREAMPlace, with comparisons against EGGPlacer and its thermal-only variant.
  • Paired statistical analyses with 95% confidence intervals for peak temperature and routed wirelength across 20 designs and three random seeds.
  • Further clarification of our methodological contribution, particularly the integration of analytical and differentiable thermal feedback throughout diffusion sampling.

If you have a moment, we would be very grateful if you could review our responses and let us know whether any concerns remain that we could further clarify.

We sincerely appreciate the time and effort you have invested in reviewing our work, and thank you for considering our additional experiments and analyses in your evaluation.

Best regards,

The Authors

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

This paper presents EGGPlacer, a diffusion-based macro placer that uses online analytical feedback and thermal guidance to account for standard-cell connectivity, space requirements, and temperature during generation. On an open-source benchmark, it reduces average routed wirelength by 2.18% relative to the strongest external baseline for this metric and achieves a lower average peak temperature under the reported thermal settings. Ablations support the contributions of both guidance terms.

Critical strengths
  1. Standard-cell feedback. Online analytical placement exposes standard-cell connectivity and space requirements to the macro generator, addressing information missing from macro-only wirelength objectives. Downstream results and ablations support its effectiveness.
  2. Joint physical guidance. The method combines wirelength and thermal objectives without retraining the diffusion model. It improves downstream results, with temperature gains independently verified by HotSpot.
Critical weaknesses
  1. Cross-design generalization. The thermal dataset covers all evaluation designs, so performance on unseen designs remains untested. A design-disjoint evaluation or a clear description of per-design calibration and its cost is needed.
  2. Thermal settings. The tenfold macro-power scaling and water cooling need stronger justification. The authors should explain the intended application and provide references, physical reasoning, or additional experiments to support these choices.
Questions for authors
  1. How well do temperature prediction and thermal guidance work on chip designs excluded entirely from thermal-model training? If a new design requires separate training or calibration, how many thermal simulations are needed, and what is the total preparation time?
  2. What is the basis for combining tenfold macro-power scaling with water cooling? Which application does this setup represent, and what references or physical analysis support it? Do the main conclusions hold with the original macro power or different cooling conditions?
4 comments
AuthorsOpenReview ↗

Response 1 to Reviewer AsWA

Dear Reviewer AsWA,

Thank you for your constructive feedback on cross-design generalization and the thermal evaluation setup. We will upload a revised manuscript with all revisions highlighted in blue for ease of reference.

W1

Cross-design generalization. The thermal dataset covers all evaluation designs, so performance on unseen designs remains untested. A design-disjoint evaluation or a clear description of per-design calibration and its cost is needed.

Response

We thank the reviewer for highlighting the need to evaluate the thermal surrogate on unseen designs. We acknowledge that our thermal surrogate is approximate and that its accuracy and validation could be further improved. However, developing a high-fidelity thermal simulator is not our primary contribution. Instead, our key contribution is to incorporate analytical and differentiable thermal feedback into the reverse diffusion process to guide macro placement generation, an approach that, to the best of our knowledge, has not been explored in prior diffusion-based macro placement methods. Moreover, developing accurate and efficient thermal simulators is itself an active and independent research direction, as demonstrated by DeepOHeat [1], ARO [2], and DeepOHeat-v1 [3]. These studies highlight ongoing challenges in prediction accuracy, computational efficiency, generalization, and reliability during optimization.

To address your request for design-disjoint evaluation, we have conducted additional hold-out experiments as detailed below. We completely exclude all samples from bp_multi57 and VeriGPU during thermal surrogate training. On the 60 test samples of bp_multi57 and 66 test samples of VeriGPU, the held-out surrogate achieves peak temperature prediction MAEs of 4.166°C and 3.315°C, respectively. For reference, the original thermal model yields MAEs of 0.445°C and 0.429°C on the same sample sets, which include samples from its training set. Its MAE on the original evaluation dataset is 0.832°C.

We further integrate this design-held-out thermal surrogate into our joint placement guidance pipeline, with all final thermal results validated via independent HotSpot evaluation. Compared with the wot (A-only) baseline variant that uses only analytical guidance, the average peak temperature across three random seeds decreases by 0.637°C on bp_multi57 but increases by 0.273°C on VeriGPU. When compared with the full EGGPlacer pipeline using the original in-distribution thermal surrogate, the held-out version delivers an average peak temperature reduction of 0.167°C on bp_multi57 and an increase of 1.813°C on VeriGPU.

These results indicate that thermal guidance can still deliver placement optimization gains on some unseen designs, but the cross-design generalization of both prediction accuracy and guidance effectiveness remains notably limited — a challenge widely recognized in the thermal modeling community. Improving generalization across a wider range of chip designs will be an important direction for future research.

Please refer to our response to Q1 for details on the training cost and data preparation overhead of the thermal surrogate.

[1] Z. Liu et al. DeepOHeat: Operator Learning-based Ultra-fast Thermal Simulation in 3D-IC Design. 2023 60th ACM/IEEE Design Automation Conference (DAC), pp. 1–6, 2023.

[2] M. Wang et al. ARO: Autoregressive Operator Learning for Transferable and Multi-fidelity 3D-IC Thermal Analysis With Active Learning. Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design (ICCAD), pp. 1–9, 2024.

[3] X. Yu et al. DeepOHeat-v1: Efficient Operator Learning for Fast and Trustworthy Thermal Simulation and Optimization in 3D-IC Design. IEEE Transactions on Components, Packaging and Manufacturing Technology, 16(1):82–96, 2026.

AuthorsOpenReview ↗

Response 2 to Reviewer AsWA

W2

Thermal settings. The tenfold macro-power scaling and water cooling need stronger justification. The authors should explain the intended application and provide references, physical reasoning, or additional experiments to support these choices.

Response

We thank the reviewer for requesting clarification of the thermal configuration and macro-power scaling.

First, we acknowledge an error in the reported thermal configuration. Upon reexamining our experimental setup, we found that the custom material configuration was not applied during HotSpot simulation. Instead, the reported results were obtained using the default material settings, comprising an active silicon die, a thermal interface material (TIM), an aluminum heat spreader, and an aluminum heat sink, without the previously described water cooling. We will correct this description and clarify the effective thermal settings in the revised paper.

Second, the 10-fold macro-power scaling was intended to evaluate thermal-aware placement under high power density conditions. The die areas of the 20 evaluated designs range from 0.2740 to 2.7989 mm², with a mean of 1.3162 mm². After scaling, their chip-level average power densities range from 0.543 to 3.540 W/mm², with a mean of 1.789 W/mm². Such high power densities are relevant to modern high-performance computing systems. For example, ATPlace2.5D [4] adopts chiplet power densities of 0.2–3.0 W/mm² in heterogeneous 2.5D benchmarks, including systems composed of CPUs, GPUs, HBMs, and DRAMs. Moreover, Yu and Zhu [5] experimentally demonstrated an average chip-level heat flux exceeding 545 W/cm² in a high-power GaN device. These studies support the relevance of our power settings for evaluating thermal-aware placement under high power density conditions. We will clarify this motivation and update the corresponding experimental settings in the revised paper.

[4] Q. Wang et al. ATPlace2.5D: Analytical Thermal-Aware Chiplet Placement Framework for Large-Scale 2.5D-IC. Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design (ICCAD), pp. 1–9, 2024.

[5] M. Yu and J. Zhu. Hotspot thermal management of silicon-based high power hetero-integration. International Journal of Heat and Mass Transfer, 203:123790, 2023.

AuthorsOpenReview ↗

Response 3 to Reviewer AsWA

Q1

How well do temperature prediction and thermal guidance work on chip designs excluded entirely from thermal-model training? If a new design requires separate training or calibration, how many thermal simulations are needed, and what is the total preparation time?

Response

We appreciate the request to clarify generalization to new designs and the preparation cost of the thermal surrogate. We refer the reviewer to our response to W1 for a discussion of thermal prediction accuracy and cross-design generalization. Below is our response regarding training.

Our thermal surrogate contains only 655,684 trainable parameters and takes approximately 6 minutes to train for 128 epochs (with early stopping) on an NVIDIA RTX 2080 Ti GPU. Thus, surrogate training itself incurs relatively low computational overhead. As detailed in Appendix F.1, we use approximately 10 base placements per design. The main preparation cost comes from the downstream OpenROAD flow (ORFS), which takes approximately 1.22 hours per placement on an Intel Xeon Gold 6226R CPU, depending on design complexity and hardware. In comparison, each HotSpot simulation takes only 56.42 seconds on average. Assuming sequential execution (each base placement is augmented with 5 noised variants requiring separate HotSpot simulations, all overhead accounted for), the total preparation time is approximately 13.2 hours per design, including ORFS, HotSpot simulation, and surrogate training.

This substantial cost of downstream physical evaluation is one of the motivations for introducing a differentiable thermal surrogate, which enables thermal feedback during diffusion sampling without repeatedly executing the full physical design and thermal simulation flow.

Q2

What is the basis for combining tenfold macro-power scaling with water cooling? Which application does this setup represent, and what references or physical analysis support it? Do the main conclusions hold with the original macro power or different cooling conditions?

Response

We thank the reviewer for asking about the physical motivation and robustness of our thermal settings. For the rationale, representative application scenarios, and supporting physical analysis of the tenfold macro-power scaling, please refer to our detailed response to comment W2.

We have not evaluated the complete pipeline under the original macro power settings or alternative cooling conditions. Therefore, the generalizability of the reported thermal improvements to these configurations remains unverified. We further note that changes in cooling conditions, material properties, or macro power levels may require separate training or calibration of the thermal surrogate.

AuthorsOpenReview ↗

Kind Follow-up on Our Rebuttal Submission

Dear Reviewer AsWA,

We hope you are doing well. We wanted to kindly let you know that we have submitted our rebuttal addressing your questions and concerns in detail, including:

  • Additional design-held-out experiments evaluating thermal surrogate generalization and guidance effectiveness on unseen designs, along with a discussion of their limitations.
  • Clarifications and corrections to the thermal simulation setup, including the effective cooling configuration and the motivation for 10-fold macro-power scaling.
  • A detailed analysis of thermal surrogate training and data preparation costs, including the required thermal simulations and downstream physical evaluation overhead.

If you have a moment, we would be very grateful if you could review our responses and let us know whether any concerns remain that we could further clarify.

We sincerely appreciate the time and effort you have invested in reviewing our work, and thank you for considering our additional evidence and clarifications in your evaluation.

Best regards,

The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Major Contributions, Strengths, and Responses to Concerns

Dear Area Chair and Reviewers,

We sincerely thank all three reviewers for their constructive feedback and recognition of our work. We have conducted additional experiments and analyses to address the concerns raised during review. Below, we summarize our core contributions, the strengths acknowledged by the reviewers, and the key rebuttal updates.

1. Core Contributions

The central goal of EGGPlacer is to bridge the gap between the objectives available during macro placement generation and the physical metrics evaluated in downstream design flows. Our contributions are threefold:

  • Bridging the objective visibility gap: We identify the limitations of macro-only HPWL (mHPWL) in learning-based placement and introduce an energy-guided diffusion framework that exposes complementary analytical and thermal feedback to macro generation.

  • Unified analytical and thermal guidance: We integrate an analytical placer as an online physical response model to expose standard-cell distribution, full netlist wirelength, and spatial competition, alongside differentiable thermal feedback. These objectives are formulated as composable energies and incorporated throughout reverse diffusion without retraining the diffusion backbone.

  • Understanding downstream wirelength proxies: We empirically demonstrate that the fidelity of HPWL as a proxy for routed wirelength varies across placement engines, highlighting the importance of aligning macro placement generation with downstream physical evaluation.

2. Acknowledged Strengths

We appreciate that the reviewers explicitly recognized several aspects of our contributions:

  • Standard-cell-aware analytical guidance (AsWA): Online analytical placement incorporates full-netlist connectivity and dynamically captures standard-cell distributions and spatial demand that are not represented by macro-only objectives.
  • Practical guidance with a frozen diffusion backbone (AsWA, wT1b, RSpk): The framework combines physical objectives without retraining, with controlled ablations and independent HotSpot evaluation supporting the analysis.
  • Comprehensive downstream evaluation (wT1b, RSpk): Evaluation through a unified ORFS and HotSpot flow on 20 public benchmarks extends beyond macro-only HPWL. In particular, RSpk recognized the value of analyzing discrepancies between intermediate wirelength proxies and final routed wirelength.

3. Key Rebuttal Updates

To address the reviewers' concerns, we have substantially expanded our experimental evaluation:

  • Thermal-aware refinement comparison (wT1b, RSpk): We added a DREAMPlace-initialized thermal-aware refinement baseline in Appendix E.4. Compared with this baseline, the thermal-only (woa) variant achieves a lower mean peak temperature, while the complete EGGPlacer achieves lower mean values for both peak temperature and routed wirelength.

  • Multi-seed evaluations and statistical analysis (wT1b, RSpk): We expanded baseline evaluations to three random seeds and conducted paired comparisons with 95% confidence intervals across 20 designs. EGGPlacer achieves updated mean ranks of 2.650 for routed wirelength and 1.600 for peak temperature. We further quantify the marginal thermal benefit and wirelength tradeoff of joint guidance.

  • Cross-design thermal generalization (AsWA, RSpk): We conducted design-held-out experiments on bp_multi57 and VeriGPU, evaluating surrogate prediction and downstream guidance effectiveness. The results reveal limitations in cross-design generalization, which we explicitly acknowledge.

  • Additional downstream physical metrics (RSpk): We supplemented the evaluation with WNS, TNS, DRC violations, and routing overflow across 20 benchmarks, providing a broader assessment of timing and routability, even though we do not design dedicated energy terms for metrics such as TNS.

  • Thermal configuration and experimental transparency (AsWA, RSpk): We corrected the reported HotSpot configuration, provided physical motivation for tenfold macro-power scaling, quantified surrogate preparation costs, and clarified the hyperparameter selection protocol.

4. Concluding Remarks

We emphasize that EGGPlacer's central contribution is a unified energy-guided diffusion framework that incorporates complementary analytical and thermal feedback into macro placement generation, rather than the development of a high-fidelity thermal simulator. The additional experiments strengthen the evaluation of this framework while clarifying the limitations of thermal surrogate generalization and the marginal benefits of thermal guidance.

We sincerely appreciate the reviewers' constructive suggestions and hope that these additional results and clarifications will be considered in the final assessment.

Best regards,
The Authors

Dear Reviewers,

As the discussion phase draws to a close, please acknowledge the author's response and let them know whether it has addressed some or all of your concerns.

Thank you for the time and effort you and the other reviewers have put into this discussion.

Best,

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.