acceptodds
Under review as a conference paper at ICLR 2027

TP3D: Timing-Aware Die-To-Die 3D Macro Placement

Abstract

Macro placement is one of the most critical steps in chip design, exerting a significant impact on power, performance, and area (PPA), while existing works primarily focus on optimizing computationally efficient surrogate objectives, resulting in a marked gap in performance improvement. The emergence of 3D integrated circuits (3D-ICs) with stacked layers connected via hybrid bonding terminals (HBTs) further complicates accurate timing estimation at early stages of the design flow. To bridge the gap, we formulate the timing-aware macro placement in 3D-ICs design as a learning-based optimization problem, and propose TP3D, a novel learning-based, timing-aware 3D macro placement framework. Specifically, TP3D introduces a structured timing prediction model that decomposes intra-die interactions and bridge-point-based inter-die interactions. The predicted metrics are further integrated into the placement procedure in the form of timing masks to guide optimization. Extensive experiments demonstrate that TP3D outperforms state-of-the-art methods, achieving an average improvement of 10.8% in Worst Negative Slack (WNS) and 8.7% in Total Negative Slack (TNS), which are of great significance for enhancing chip performance. Notably, this work highlights the importance of integrating structured, learnable objective modeling for intractable objectives, compared to directly optimizing surrogate objectives.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Electronic Design Automation, Macro placement

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_pXfDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TP3D predicts WNS/TNS from macro placements with separate GNNs for intra-die and inter-die interactions; cross-die paths use latent bridge points. Predicted timing costs guide placement masks and evolutionary refinement. Trained on 1,000 placements from five designs, it is evaluated on seven Open3DBench designs, including two held out. The authors report 10.8% mean WNS and 8.7% TNS improvement over four baselines, with Kendall correlations of 0.606/0.704 for predicted WNS/TNS.

Critical strengths
  1. The intra-die/inter-die model and latent bridge points reflect distinct 3D path structures; the paper shows HPWL correlates poorly with timing.
  2. Full-flow timing improves on seven designs, including two held out; TP3D improves WNS on all seven.
Critical weaknesses
  1. Predictor validation is limited. Reported Kendall correlations are 0.606/0.704 for pWNS/pTNS and 0.567/0.534 for the combined score. The evaluation split is unclear, held-out prediction accuracy is absent, and three runs may not characterize variation from random initialization and swaps.
  2. Scope and representation are narrow. The method targets two-die F2F bonding. Its macro-only graph summarizes standard-cell paths, but no ablation tests this abstraction.
  3. There is no explanation of how the authors handle the variability from macro/gloabel placer across different runs. The layouts produced by OpenRoad can be very different with different configurations.
Questions for authors

See the weaknesses above.

4 comments
AuthorsOpenReview ↗

Response to Reviewer pXfD

Response to Reviewer pXfD

We thank the reviewer for this careful audit. We answer each weakness and question in order.

W1: Predictor validation, split and held-out accuracy

The paper does not state the evaluation split clearly, and that was our presentation failure. The setting is five training designs (ariane133, ariane136, bp_be, bp_fe, bp_multi) with 200 placements each, split 80/20 per design; bp_quad and swerv_wrapper are excluded from training, from the normalization statistics and from checkpoint selection. The reported 0.606/0.704 for pWNS/pTNS and 0.567/0.534 for the combined score are test-split correlations.

What we did not state, and should have, is the aggregation. A correlation pooled across designs with different label scales can be inflated by design identity, and it is not the same quantity as within-design candidate ranking, which is what the placement procedure actually needs.

Measured per-design correlations. Kendall τ is computed separately inside each design on that design's test placements, so design identity enters no correlation and the quantity is the one the placement search relies on — ordering candidates within one design. The values are given in Table B.

On the three-run concern: we agree that three runs do not characterise initialization and swap variation. The controls now use five macro seeds per design.

W2: Two-die F2F scope and the graph-abstraction ablation

On the scope. The two-die face-to-face setting is what we measured and all reported numbers are for that setting. The formulation itself is not restricted to two dies: the intra-die and inter-die terms are computed by two separate GNNs, one for macro relations within a die and one for relations across dies. Extending to nn dies requires adding the corresponding inter-die relation for each die pair, rather than changing the model structure, and the paired masks generalise in the same way. We have not measured multi-die or TSV-based stacking, and we will state the scope in those terms rather than claiming results we do not have.

On the abstraction ablation. The macro graph, which summarises standard-cell paths into macro-to-macro relations through path-depth features instead of modelling cells as graph nodes, is taken from LaMPlace, and the appendix states this inheritance. We do not claim it as a contribution of this paper, so the ablation the reviewer asks for would test an inherited component rather than anything we introduce. For that reason we have not run it, and we will make the attribution explicit in the text. Ablating the abstraction, for example by comparing the macro-level graph against a cell-level or path-level representation, is a reasonable next step and we will list it as future work.

W3: Run-to-run variation and Open3DBench configurations

Same configuration for every method. Back-end evaluation is carried out with Open3DBench. All methods we compare share one fixed Open3DBench configuration within a design: the same input netlist, clock period taken from the distributed SDC, the same PDK (nangate45), the same standard-cell placement stage (Analytical-M3D), and the same timing extraction stage. The only input that differs between methods is the macro placement file itself. Because no configuration parameter is changed between methods, the improvement cannot come from a different back-end setup.

Table A: TP3D per-design WNS/TNS over five macro seeds

Format: WNS / TNS in ns, mean ± standard deviation; closer to zero is better.

DesignWNSTNS
ariane133−0.13 ± 0.01−51.88 ± 7.57
ariane136−1.09 ± 0.02−2597.06 ± 34.28
bp_be−0.31 ± 0.01−28.41 ± 0.42
bp_fe−0.85 ± 0.01−83.72 ± 1.86
bp_multi−5.36 ± 0.01−2844.26 ± 67.18
bp_quad−0.57 ± 0.04−5263.58 ± 347.59
swerv_wrapper−0.39 ± 0.03−134.86 ± 25.56

Table B: Per-design Kendall τ

Kendall τ is computed within each design on that design's test placements. swerv_wrapper and bp_quad are held out of training.

DesignpWNS τpTNS τ
ariane1330.52230.6976
ariane1360.56210.6691
bp_be0.68070.5883
bp_fe0.53120.6529
bp_multi0.51040.6148
bp_quad (held-out)0.54750.6285
swerv_wrapper (held-out)0.59540.6437

Dear Reviewer pXfD,

We are the authors of the paper "TP3D: Timing-Aware Die-To-Die 3D Macro Placement". Thank you very much for the valuable comments you provided during the review. We would appreciate it if you could review our responses when you have a chance. We hope they address your concerns and clarify our work. If you find that your concerns have been sufficiently addressed, we would be grateful if you could consider raising your score. If any questions or concerns remain, please let us know. We would be very grateful if you could spare the time to reply.

Best regards, The Authors

~Daying_Sun1OpenReview ↗

Thanks for the rely

Thanks for the effort from all authors to address my questions and concerns. I have no more questions. And I would like to retain my score.

AuthorsOpenReview ↗

Some questions about public comment

Dear Reviewer pXfD,

We just received a response from the public comment. The comment is "Thanks for the effort from all authors to address my questions and concerns. I have no more questions. And I would like to retain my score." We are confused whether this is from the reviewer. If so, since we have addressed all your questions, we still hope you can consider raising your score. We would really appreciate it if you could. If not, we still hope to get your reply.

Best regards, The Authors

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

TP3D addresses timing-aware macro placement for face-to-face 3D integrated circuits. It learns an additive surrogate for downstream WNS and TNS from macro graphs and positions, distinguishing intra-die distances from inter-die interactions parameterized by latent bridge points (LBPs). The surrogate generates two placement masks, used with legal-position masks, greedy placement, and evolutionary search. Training uses 1,000 placements from five Open3DBench designs; evaluation covers seven designs, including two excluded from training. Table 1 improves mean WNS over the selected baselines on every design, while TNS gains vary across designs.

Critical strengths
  1. The problem and evaluation target are meaningful. Optimizing early placement using expensive downstream timing feedback addresses a real limitation of wirelength objectives. The common downstream flow and three independent placement runs in Table 1 are valuable. The approximately 10.8% average WNS improvement is supported by the displayed results, including improvements on both unseen designs. This is relevant application-oriented ML evidence, even without a new general-purpose GNN architecture.

  2. The 3D inductive bias preserves an efficient optimization interface. Separate intra/inter-die interactions and learned two-segment geometry retain the additive structure needed to generate masks. Figure 6 supports the LBP parameterization: removing it reduces the reported WNS/TNS Kendall correlations from 0.606/0.704 to 0.362/0.447. This is a substantive ablation, although its downstream implications remain incompletely established.

Critical weaknesses
  1. The novel contribution needs a sharper boundary and stronger validation. LaMPlace, ICLR 2025 already combines symmetric GNN-derived Laurent coefficients, regression and ranking losses, learned placement masks, and greedy/evolutionary search. TP3D cites it in Appendix B.1, but the principal method discussion does not explain this extensive relationship. The real increment is the die-conditioned model and LBP geometry. The existing no-LBP experiment is relevant; please specify what it retains and connect its ranking improvement to final placement quality. Related-work positioning should also acknowledge timing-driven 3D placement, including the cited TA3D and Ji et al., DAC 2026. The latter has different physical settings and mixed-size optimization, so its published numbers are not directly comparable; compatibility should be explained rather than assumed.

  2. The ranking evidence does not yet establish reliable transfer for the optimization task. Sections 5.3–5.4 do not clearly identify the split or aggregation behind Figures 5–6. Ranking different designs by timing scale can differ substantially from ranking placements within one design. Only two designs test transfer, and their TNS outcomes differ: swerv_wrapper improves, whereas bp_quad worsens from −4097.09 to −5250.15 ns, a 28.1% increase in violation magnitude. The WNS improvement remains valid, but the overall timing trade-off and generalization claim require qualification. These limitations matter more than a request to expand into unrelated thermal or power objectives.

Questions for authors
  1. For the no-LBP variant, are separate intra/inter-die predictors retained? Which parameters change, and are downstream WNS/TNS results available under the same placement/search settings? Please explicitly distinguish the inherited LaMPlace components from the new 3D contribution.

  2. Which samples generate Figures 5–6? Please provide within-design correlations, including the two unseen designs if predictions exist, and specify whether ranking pairs cross design boundaries and how labels are normalized.

  3. Table 1 treats higher signed slack as better, whereas Algorithm 2 minimizes predictions. What sign and scale transformations are applied? How are zero distances and self-pairs handled for negative exponents, including legal coincident horizontal coordinates on opposite dies? These are implementation clarifications, not established bugs.

  4. Please clarify the average TNS improvement: averaging per-design relative changes against Analytical-M3D from the displayed means gives about 8.4%, versus the stated 8.7%. Also reconcile the five node features and edge dimension in Appendix B.1 with the input dimensions in Table 4.

6 comments
AuthorsOpenReview ↗

Response to Reviewer rJrB

Response to Reviewer rJrB

We thank the reviewer for this careful audit. We answer each weakness and question in order.

W1 + Q1: Inherited components, no-LBP and same-search downstream results

We agree fully. We will name the inherited parts in the related-work section (Laurent coefficients, GNN, macro-level graph, pairwise ranking loss, learned-mask WireMask-EA). Our contribution is die-conditioned separate intra/inter predictors, a learned two-segment bridge point, and paired timing-masks that select a legal (x,y,z)(x, y, z).

What no-LBP retains. It keeps both the intra-die and inter-die GNNs and both sets of coefficient matrices; the only change is that the inter-die distance term goes from two segments to one direct distance.

Same-search downstream WNS/TNS. Results are in Table A at the end of this response.

Removing LBP degrades both WNS and TNS on all seven designs, which shows that the direction sensitivity of the inter-die term does have a downstream effect. On the ranking loss: the paper's Figure 3 already shows that the MSE-only variant without ranking loss fails to converge stably, with the unstable ranges marked in the figure. Since that model does not converge, using its output for placement and simulation would measure training failure rather than the function of the loss term; we therefore did not add a downstream measurement for it.

TA3D and Ji et al. We reproduced TA3D and use it as a baseline; the comparison is in Table A.

LaMPlace-3D baseline. We built a per-die LaMPlace: the design is first split into a top and a bottom die using TA3D's partitioning method, then LaMPlace is run separately on each die. The models were retrained on the five training designs, applied die by die, and passed through the same back-end flow. The comparison is in Table A.

W2 + Q2: Split, aggregation, transfer and correlation construction

The dataset is 200 placements from each of five designs, split 80/20 per design, with bp_quad and swerv_wrapper excluded from training, normalization and checkpoint selection. Ranking pairs are constructed within a design — the reviewer is right that cross-design pairs would mix label scales, and that ranking designs by scale is not ranking placements within a design. We will state this explicitly.

On transfer, we have stated both outcomes in the text rather than presenting two tests as broad generalization: bp_quad TNS worsens by 28.1% while WNS improves, and swerv_wrapper improves in both (Table A).

Measured per-design correlations. Kendall τ is computed separately inside each design on that design's test placements, so no pair crosses a design boundary and no correlation can be inflated by design identity — the concern the reviewer raises. The values are given in Table B.

Q3: Sign/scale transform, zero distance and self-pairs

Thank you for framing these as implementation clarifications rather than established bugs; we read them the same way.

Sign and scale. For signed slack ss the training target is y=(−s−μ)/σy = (-s-\mu)/\sigma. A larger slack gives a smaller target cost, and MSE, ranking and the search minimization all use yy consistently.

Zero distance. No single epsilon: all inverse distances are capped at 100 and same-cell pairs are zeroed. Because txt^x and tyt^y are sigmoids in (0,1)(0,1), the bridge lies strictly inside the box between the two macros, so rib+rjb>0r_{ib} + r_{jb} > 0 for distinct macros and the singular terms never see a zero argument — including legal cross-die coincidence, which only makes one coordinate difference zero while the sum of the two segment lengths remains positive.

Q4: 8.4% versus 8.7%, and node/edge dimensions

The average is 8.40%; the 8.7% in the text is a typo we will correct. Appendix B.1 and Table 4 describe different objects: B.1 defines the graph input (five node channels; de=D+D′+4=14d_e = D + D' + 4 = 14 edge channels at D=9D = 9, D′=1D' = 1), while Table 4 lists first-layer encoder weight shapes (3×2563 \times 256 node, 16×25616 \times 256 edge). We will add a sentence distinguishing them.

AuthorsOpenReview ↗

More Table Results

Table A: Per-design WNS/TNS for the requested controls

Format: WNS / TNS in ns; closer to zero is better.

DesignTP3Dno-LBPTA3DLaMPlace-3D
ariane133−0.13 / −52.02−0.255 / −78.60−0.160 / −62.06−0.19 / −91.30
ariane136−1.07 / −2602.25−2.02 / −4737.18−1.724 / −3357.79−1.91 / −2751.8
bp_be−0.32 / −27.69−0.949 / −126.2−0.47 / −59.51−0.46 / −74.0
bp_fe−0.85 / −84.83−1.031 / −104.15−0.940 / −97.92−1.602 / −92.6
bp_multi−5.37 / −2837.15−6.264 / −4376.22−7.170 / −4722.13−6.613 / −2966.0
bp_quad−0.59 / −5250.15−0.68 / −7742.14−0.68 / −7742.14−1.88 / −11792.5
swerv_wrapper−0.39 / −134.75−0.642 / −261.9−0.606 / −219.78−0.87 / −793.9

Table B: Per-design Kendall τ

Kendall τ is computed within each design on that design's test placements. swerv_wrapper and bp_quad are held out of training.

DesignpWNS τpTNS τ
ariane1330.52230.6976
ariane1360.56210.6691
bp_be0.68070.5883
bp_fe0.53120.6529
bp_multi0.51040.6148
bp_quad (held-out)0.54750.6285
swerv_wrapper (held-out)0.59540.6437

Dear Reviewer rJrB,

We are the authors of the paper "TP3D: Timing-Aware Die-To-Die 3D Macro Placement". Thank you very much for the valuable comments you provided during the review. We would appreciate it if you could review our responses when you have a chance. We hope they address your concerns and clarify our work. If you find that your concerns have been sufficiently addressed, we would be grateful if you could consider raising your score. If any questions or concerns remain, please let us know. We would be very grateful if you could spare the time to reply.

Best regards, The Authors

Reviewer_cug3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3(+1)
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses timing-driven macro placement for face-to-face (F2F) hybrid-bonded two-die 3D ICs, motivated by the weak correlation it reports between HPWL and post-route WNS/TNS (Fig. 1). TP3D learns an offline predictor that decomposes WNS/TNS into pairwise macro terms: a Laurent polynomial of horizontal distance for same-die pairs and, for cross-die pairs, a two-segment distance through a learned "latent bridge point" (LBP). The coefficients come from GNNs over a macro-level netlist graph, trained with MSE plus a pairwise ranking loss (Sec. 4.1–4.2).

The trained predictor is converted into top- and bottom-die grid "timing masks" that drive a greedy, WireMask-EA-style evolutionary placer (Sec. 4.3, Alg. 1–2). On seven Open3DBench designs (two held out from training) evaluated through an OpenROAD-based 3D flow, TP3D obtains the best mean WNS on all designs (reported average gain 10.8%) and the best TNS on five of seven (reported 8.7%; Table 1).

Critical strengths

S1: Relevant problem, practical two-die design, end-to-end evaluation. Timing-driven macro placement for F2F 3D ICs is important, and the paper evaluates with a full post-route 3D flow (Open3DBench/OpenROAD) rather than a surrogate. The two-die design (die-gated coefficients in Eq. 9, per-die masks and die selection in Alg. 2) is simple, fast at inference (Table 5) and comes with an anonymous code link.

S2: Consistent WNS gains over the strongest baseline. TP3D has the best mean WNS on all seven designs against Analytical-M3D (Table 1; sign test p ≈ 0.016), with the largest gains on the held-out bp_quad (−0.74 → −0.59 ns) and swerv_wrapper (−0.53 → −0.39 ns). Its TNS gains on ariane136, bp_be and bp_fe are significant (Welch p ≤ 0.01), and Fig. 6 and App. C.3–C.5 add predictor ablations and sensitivity studies.

Critical weaknesses

W1: The core framework is LaMPlace's, without attribution or comparison. The Laurent-polynomial predictor with bilinear GNN coefficients (Eqs. 1–2, same K), the graph features (App. B.1), the weighted ranking loss (Sec. 4.2, App. B.7), the learned mask substituted into the WireMask-EA placer (Alg. 1–2) and the periphery case study (Sec. 5.5) all follow LaMPlace (ICLR 2025), cited only in App. B.1. Please frame TP3D as its 3D extension, attribute these parts, and compare with a LaMPlace-style baseline.

W2: The LBP is weakly motivated and lacks end-to-end evidence. Whenever t^x = t^y, the LBP term is just a rescaled r_ij^k (Eqs. 6–7), and under rectilinear routing every bridge point in the bounding box gives the same length, so Fig. 3's argument does not by itself explain a timing difference. Negative exponents also diverge for vertically aligned cross-die macros. Evidence is only predictor-level (Fig. 6); please compare with Manhattan or fixed-midpoint distances end-to-end.

W3: Baselines and controls cannot isolate the gains. The four baselines (Sec. 5.1) lack a timing-driven method (e.g., TA3D), a classical placer (e.g., Hier-RTLMP) and a periphery heuristic, although TP3D mainly moves macros to the periphery (Fig. 7). Also missing: the same search with an HPWL wiremask, and best-of-200 from the labeled Analytical-M3D data. Baseline values differ greatly from Open3DBench v1 (ariane133 Open3D-Tiling TNS −161.67 vs. −3049.41 ns); please state clock and flow settings.

W4: The correlation evidence is under-specified and possibly confounded. Eq. 11 pairs samples across designs, and Figs. 1 and 5 do not say whether τ is pooled; with labels differing ~100× across designs and ~80% of validation pairs cross-design, recognizing the design alone can inflate τ. HPWL–TNS τ = −0.331 is in the expected direction and exceeds pWNS–TNS (0.306), so "almost uninformative" (Sec. 3) is overstated. Please report per-design τ, also on held-out designs and TP3D outputs.

W5: The gains are mixed, often not significant, and narrower than claimed. TNS is worse than Analytical-M3D on bp_multi (−12.8%) and the held-out bp_quad (−28.1%), and the abstract's 8.7% TNS average does not reproduce from Table 1 (8.4%). With three runs, Welch's t-test finds the WNS gain significant only on bp_multi (held-out designs: p ≈ 0.09 and 0.17), and "zero-shot" rests on two held-out designs. More seeds, significance tests and diverse held-out designs are needed.

Questions for authors
  1. Relation to LaMPlace. Beyond the LBP, the intra/inter-die split, the two-die masks and the loss clipping, which components differ from LaMPlace? Please report end-to-end WNS/TNS for a LaMPlace-style 3D baseline under the same search budget (e.g., TP3D without LBP and die-specific coefficients).

  2. End-to-end ablations and controls. Please report final WNS/TNS for (i) TP3D without LBP, (ii) the LBP replaced by Manhattan distance or a fixed midpoint, (iii) the same greedy + EA search driven by an HPWL wiremask, and (iv) the best of the 200 labeled Analytical-M3D placements per training design.

  3. Computation of Kendall's τ. Is τ computed within each design or pooled across designs, and on which split? Please give per-design τ on validation data, on bp_quad and swerv_wrapper, and on TP3D-generated placements, and state the label sign convention used in Fig. 1, Fig. 5 and Alg. 2.

  4. LBP implementation. How are the k ∈ {−1, −2} terms handled as r_ib or r_jb approaches zero, e.g., at the grid cell facing a placed macro in Alg. 1? Many cross-die macros in Figs. 7(b) and 12 appear nearly stacked, about one grid cell apart; is this the singular term at work? Since the term is unchanged under t → 1 − t, how should the bridge point be interpreted physically?

  5. Evaluation settings. Which clock periods, PDK and flow settings were used, and why do the Open3D-Tiling/DMP values differ from Open3DBench's? How was the 8.7% TNS figure computed, and were any outputs selected using ground-truth timing (as in LaMPlace's best-of-five protocol)?

5 comments
AuthorsOpenReview ↗

Response to Reviewer cug3

Response to Reviewer cug3

We thank the reviewer for this careful audit. We answer each weakness and question in order.

W1 + Q1: Relation to LaMPlace, and the same-budget 3D baseline

We will state plainly in the related-work section: the GNN, the macro-level graph abstraction, the weighted ranking loss, and the insertion of learned masks into WireMask-EA are all inherited from LaMPlace and are not claimed as new. Our contribution is die-conditioned separate intra/inter predictors, a learned two-segment bridge point, and paired timing-masks that select a legal (x,y,z)(x, y, z).

LaMPlace-3D baseline. We built a per-die LaMPlace: the design is first split into a top and a bottom die using TA3D's partitioning method, then LaMPlace is run separately on each die. The models were retrained on the five training designs, applied die by die, and passed through the same back-end flow.

TP3D is better in both WNS and TNS on all seven designs (Table A).

On no-LBP. The end-to-end LBP ablation is in Table A.

W2 + Q4: LBP motivation, evidence, stabilization and physical reading

In fact, txt_x and tyt_y are computed separately. We counted this over the five training designs and found that txt_x and tyt_y differ in about 93% of macro pairs. When the two differ, the two segment lengths depend on the direction of the horizontal displacement, not only its norm.

Stabilization. All inverse distances are capped at 100 and same-cell pairs are zeroed. Because tt is a sigmoid in (0,1)(0,1), the bridge lies strictly inside the box spanned by the two macros, so rib+rjb>0r_{ib} + r_{jb} > 0 and the singular terms never receive a zero argument on a real cross-die pair.

End-to-end Manhattan and fixed-midpoint controls, rerun on the original framework. Manhattan changes only the distance in the original paper from Euclidean to Manhattan; fixed-midpoint only fixes the LBP location at the midpoint of the two horizontally separated points.

Both alternatives are worse than the full model in WNS and TNS on six of the seven designs, with fixed-midpoint degrading substantially more (Table A). The remaining case is swerv_wrapper, where Manhattan gives a slightly better WNS (−0.35 vs −0.39) and a better TNS (−132.5 vs −134.86).

Stabilization of the negative exponents (Q4). All inverse distances are capped at 100, and pairs in the same grid cell have their inverse values set to zero before the cap. Because tt is a sigmoid in (0,1)(0,1), the learned bridge lies strictly inside the box spanned by the two macros, so rib+rjb>0r_{ib} + r_{jb} > 0 for any two distinct macros and the singular k=−1,−2k = -1, -2 terms never receive a zero argument on a real cross-die pair. Near-stacking in the visualizations cannot by itself confirm or exclude a singularity artifact, so we do not use it as evidence. On the physical reading: the bridge is not a physical location but a latent reparameterization that makes the inter-die response direction-sensitive when tx≠tyt_x \neq t_y, and we will stop attributing physical meaning to it.

W3 + Q2: Controls, baselines and end-to-end ablations

Timing-aware TA3D (W3). We reproduced TA3D's method and use it as a baseline against the main method. TP3D is better in both WNS and TNS on all seven designs (Table A).

Same-search HPWL mask (W3, Q2). FlexPlanner in Table 1 is itself an HPWL-mask method: we removed the two metrics used only in the floorplan stage and retained HPWL as the macro placement objective. It is a substantially stronger HPWL comparator than a plain greedy+EA search. TP3D outperforms FlexPlanner on all seven designs in both WNS and TNS — for example ariane133 −0.13 / −51.88 versus −0.17 / −124.29, and bp_fe −0.85 / −83.72 versus −0.96 / −111.11 (Table A). The HPWL-mask category is therefore covered by a stronger comparator than the one requested.

Removing LBP (W3, Q2). Removing LBP degrades both WNS and TNS on all seven designs (Table A). This does not contradict the analysis above: LBP does not change wirelength, but it changes how sensitive the inter-die term is to the direction of horizontal displacement; without LBP that degree of freedom collapses into a single distance.

AuthorsOpenReview ↗

More Rebuttal

Periphery heuristic (W3). We do not report a separate periphery heuristic, because in our flow it is not a separable factor: die-level partitioning itself produces a periphery-style layout, since assigning macros to the periphery of each die is the natural outcome of that step. An independent edge heuristic therefore cannot isolate what the reviewer wants to test, because our per-die baseline is already an edge-style layout. The LaMPlace-3D baseline above is exactly that control: it partitions by die, produces periphery-style macro layouts, but lacks the timing-aware prediction and masks. It is worse than TP3D on all seven designs, which indicates that periphery placement alone does not reproduce TP3D's timing behaviour. We will revise Section 5.5 accordingly, presenting edge placement as a shared property of the die-level flow rather than as TP3D's core mechanism, and attributing the difference to which macros the timing masks bring together or separate.

Hier-RTLMP (W3). Hier-RTLMP is a 2D method and has empirically been surpassed by LaMPlace, which we already compare against. This paper addresses a 3D two-die problem, and adding a 2D baseline cannot isolate the source of gains in that setting. This is a scoping decision.

Best-of-200 labelled placements (W3, Q2). Selecting the best of 200 labelled placements requires 200 ground-truth timing queries, whereas TP3D at inference performs a single forward prediction. The compute and timing-query budgets are not comparable, so this is a data-reuse oracle rather than an inference baseline. We do not present it as a comparator and make no claim that TP3D beats that oracle.

The ariane133 discrepancy and flow settings (W3). The two numbers do not come from one protocol. Open3DBench v1 runs its own end-to-end flow and reports −3049.41 ns, whereas our Table 1 row for Open3D-Tiling is a macro placement result with standard cells placed by Analytical-M3D and timing taken from our common flow, hence −161.67 ns. Section 5.2 states this but does not connect it to the table row; we will add that sentence. Clock: 4.0 ns for ariane133 and 6.0 ns for ariane136, from the distributed SDC; PDK nangate45 for all designs.

W4 + Q3: Correlation split, pooling and label signs

The reviewer is right: −0.331 is in the expected direction, and we will stop calling it almost uninformative in the revision. Ranking pairs are constructed within a design, since cross-design pairs would mix label scales. Labels use y=(−s−μ)/σy = (-s-\mu)/\sigma, so a larger slack gives a smaller target and the minimization direction is consistent throughout.

Measured per-design correlations. Kendall τ is computed separately inside each design on that design's test placements, so design identity enters no correlation and the value is the quantity the placement search actually relies on — the ordering of candidates within one design. The values are given in Table B.

W5 + Q5: Effect size, significance and flow settings

Recomputed from the displayed means: WNS 10.83% and TNS 8.40%. The 8.7% in the text is a typo and will be corrected. Both regressions are stated in the text rather than only in the table: bp_multi TNS +12.84% and bp_quad TNS +28.14%. TP3D has the best mean WNS on 7/7 designs and the best mean TNS on 5/7.

We agree that three runs cannot establish per-design significance. The controls now use five macro seeds per design, and all comparison data above are five-seed means. Table A reports the five-seed means for TP3D as well, so that every comparison inside it is like-for-like; the TP3D cells there therefore differ slightly from the single-run values in Table 1 of the paper.

AuthorsOpenReview ↗

More Table Results

Table A: Per-design WNS/TNS for the requested controls

All columns report five-seed means. Every control column was rerun on the original framework under a common flow. Format: WNS / TNS in ns; closer to zero is better.

DesignTP3Dno-LBPManhattanfixed-midpointTA3DLaMPlace-3D
ariane133−0.13 / −51.88−0.255 / −78.60−0.25 / −62.18−0.337 / −3618.1−0.160 / −62.06−0.19 / −91.30
ariane136−1.09 / −2597.06−2.02 / −4737.18−1.31 / −2810.7−1.74 / −3069.55−1.724 / −3357.79−1.91 / −2751.8
bp_be−0.31 / −28.41−0.949 / −126.2−0.39 / −27.4−0.42 / −74.0−0.47 / −59.51−0.46 / −74.0
bp_fe−0.85 / −83.72−1.031 / −104.15−0.920 / −95.0−0.912 / −103.6−0.940 / −97.92−1.602 / −92.6
bp_multi−5.36 / −2844.26−6.264 / −4376.22−6.071 / −2974.1−6.184 / −3014.2−7.170 / −4722.13−6.613 / −2966.0
bp_quad−0.57 / −5263.58−0.68 / −7742.14−0.68 / −7742.14−0.92 / −10461.27−0.68 / −7742.14−1.88 / −11792.5
swerv_wrapper−0.39 / −134.86−0.642 / −261.9−0.35 / −132.5−0.637 / −165.4−0.606 / −219.78−0.87 / −793.9

Table B: Per-design Kendall τ

Kendall τ is computed within each design on that design's test placements. swerv_wrapper and bp_quad are held out of training.

DesignpWNS τpTNS τ
ariane1330.52230.6976
ariane1360.56210.6691
bp_be0.68070.5883
bp_fe0.53120.6529
bp_multi0.51040.6148
bp_quad (held-out)0.54750.6285
swerv_wrapper (held-out)0.59540.6437
Reviewer_LXtmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TP3D, a timing-aware macro placement framework for face-to-face bonded, two-die 3D integrated circuits that addresses the mismatch between surrogate objectives such as HPWL and downstream timing metrics. TP3D uses graph neural networks to model intra-die and inter-die interactions separately, introducing latent bridge points to capture inter-die connection geometry. Trained with mean squared error and pairwise ranking losses, the model generates two-layer timing masks that guide placement alongside legality masks and evolutionary search. Experiments on seven Open3DBench designs, including two unseen during training, show the best mean WNS on all seven designs and the best mean TNS on five, with reported average improvements of 10.8% and 8.7%, respectively, over Analytical-M3D. Ablation studies further support the contributions of latent bridge points and ranking loss to prediction-ranking accuracy.

Critical strengths

The problem is practically important, and evaluation uses downstream timing results. The paper evaluates WNS/TNS after subsequent standard-cell placement and the Open3DBench flow, rather than relying solely on HPWL reduction. In Table 1, TP3D achieves the best mean WNS on all seven designs and the best mean TNS on five. These results provide meaningful empirical support for using learned timing objectives during early macro placement.

The structured prediction model can be integrated into placement search. Modeling intra-die and inter-die interactions separately, then converting pairwise costs into two-layer position masks, allows the predictor to guide candidate-position selection directly. In Figure 6, removing LBP modeling reduces the WNS/TNS Kendall correlations from 0.606/0.704 to 0.362/0.447, providing initial evidence for the value of the inter-die modeling.

Critical weaknesses

There are unresolved consistency issues between the objective function and the optimization algorithm. The paper uses actual WNS/TNS values as prediction labels, and the ranking loss encourages predictions to follow the same ordering as those labels. However, Algorithms 1 and 2 select minimum timing-mask and prediction values, while Table 1 explicitly states that higher WNS/TNS values are better. If the implementation negates slack or converts it into a violation measure, the complete definition should be provided. Otherwise, readers cannot verify that the algorithm optimizes in the intended direction. Moreover, the default exponent set includes −1 and −2, while macros on different dies may share identical horizontal coordinates, potentially producing zero distances raised to negative powers. The paper does not specify distance clipping or another stabilization mechanism. These issues directly affect whether the objective is well defined and the algorithm reproducible.

Questions for authors

Q1: Are the training labels negated, normalized, or otherwise transformed? Please specify the complete mapping from raw WNS/TNS to training targets and then to the minimized placement cost, and explain how the ranking loss is consistent with minimization in Algorithm 2.

Q2: How are (r^{-1}) and (r^{-2}) handled when macros on different dies share the same horizontal coordinates, or when a candidate grid point coincides horizontally with an already placed macro? Do you use an epsilon, a minimum distance, or candidate exclusion? Please provide the exact settings.

Q3: Which data split is used to compute the Kendall correlations in Figure 5? Are correlations calculated separately for each design and then aggregated, or computed after pooling samples across designs? Please provide per-design correlations for the validation set and for bp_quad and swerv_wrapper, and clarify whether results on these two unseen designs were used for hyperparameter selection.

Q4: Under identical initialization and search budgets, could you report final WNS/TNS for variants without LBP, without ranking loss, and with conventional wiremasks? These results would directly inform the assessment of each component’s necessity and contribution.

Q5: Appendix B.1 defines five-dimensional node features. Equation (17), with (D=9,D'=1), implies fourteen-dimensional edge features, whereas Table 4 lists three-dimensional node inputs and sixteen-dimensional edge inputs. Which features are actually used? In addition, the z values assigned in the top/bottom selection branches of Algorithm 2 appear reversed relative to the die-index convention in the main text. Please verify this.

3 comments
AuthorsOpenReview ↗

Response to Reviewer LXtm

Response to Reviewer LXtm

We thank the reviewer for this careful audit. We answer each weakness and question in order.

W1 + Q1: Label direction and the raw-to-target-to-cost mapping

The paper does not give the complete mapping. Our convention is as follows.

For signed slack ss, the training target is

y=−s−μσy = \frac{-s - \mu}{\sigma}

where μ\mu and σ\sigma are normalization statistics. A larger (better) ss therefore yields a smaller yy. MSE and the pairwise ranking loss both consume this same yy, and the search minimizes the predicted yy directly; no additional negation is applied at inference. The three stages are therefore consistent without any compensating sign flip. We will add this chain to Appendix B.3.

W2 + Q2: Singular exponents, zero distance and stabilization

The paper does not specify this, and we should have. The handling is not a single epsilon but three steps:

  1. All inverse distances — the direct term and both LBP segments — are capped at 100.
  2. Pairs falling in the same grid cell have their inverse values set to zero before the cap.
  3. Because tx,ty=sigmoid(⋅)∈(0,1)t^x, t^y = \mathrm{sigmoid}(\cdot) \in (0,1), the learned bridge lies strictly inside the axis-aligned box spanned by the two macros. For any two distinct macros rib+rjb>0r_{ib} + r_{jb} > 0, so the singular k=−1,−2k = -1, -2 terms never receive a zero argument on a real cross-die pair.

Training, the greedy mask and the EA scorer all use the same treatment. A candidate point horizontally coincident with a placed macro only makes one coordinate difference zero; the sum of the two segment lengths is still positive, so it is safe as well.

Q3: Split, aggregation and per-design correlations

The dataset is 200 placements from each of five training designs (1000 total), split 80/20 per design. bp_quad and swerv_wrapper are excluded from training, from the normalization statistics and from checkpoint selection, so the unseen designs do not influence hyperparameters or model selection.

Measured per-design correlations. Kendall τ is computed separately inside each design on that design's test placements, so no correlation pools placements from different designs and design identity cannot inflate any value in the table. The values are given in Table B.

Q4: Matched-budget ablations (no-LBP, no-ranking, wiremask)

We ran the same-search suite, changing only the LBP or the loss, with five macro seeds per design.

On "conventional wiremask". FlexPlanner in Table 1 is itself an HPWL-mask method: we removed the two metrics used only in the floorplan stage and kept HPWL as the macro placement objective. It is a substantially stronger HPWL comparator than a plain greedy+EA search, and TP3D outperforms it on all seven designs in both WNS and TNS — for example ariane133 −0.13 / −51.88 versus −0.17 / −124.29, and bp_fe −0.85 / −83.72 versus −0.96 / −111.11. The HPWL-mask category is therefore covered by a stronger comparator than the one requested.

On removing LBP. Removing LBP degrades both WNS and TNS on all seven designs (Table A). This is consistent with the analysis in Section 2: LBP does not change wirelength, but it changes how sensitive the inter-die term is to the direction of horizontal displacement; without it that degree of freedom collapses into a single distance.

On removing the ranking loss. This is already reported in the paper's ablation. Figure 3 shows that the MSE-only variant without the ranking loss fails to converge stably, and the figure marks the upper bounds of its unstable ranges with hatching. Since the model itself does not converge, using its output to place macros and then running timing simulation would not yield a meaningful conclusion about the role of the ranking loss: the performance of a layout produced by an unconverged predictor reflects training failure, not the function of that loss term. We therefore did not add a downstream placement measurement for this variant and instead cite the paper's existing convergence evidence — the ranking loss is what makes relative ordering learnable, whereas WNS and TNS are produced after several downstream EDA stages and direct regression of their absolute values is highly sensitive to noise.

Q5: Feature dimensions and the z assignment

These are two different objects and we should have said so. Appendix B.1 defines the graph input: five node channels (size_x, size_y, node_area, sqrt_node_node, num_pins) and de=D+D′+4=14d_e = D + D' + 4 = 14 edge channels at D=9D = 9, D′=1D' = 1. Table 4 reports first-layer encoder weight shapes (3×2563 \times 256 for the node MLP, 16×25616 \times 256 for the edge MLP), not the graph feature count. We will add a sentence separating the two.

On the z assignment: z=0z = 0 is the bottom die and z=1z = 1 the top die. Each branch places the intra-die term on the die containing the macro being placed and the inter-die term on the mask of the opposite die. We rechecked the branch conditions and the text is consistent; there is no reversal.

AuthorsOpenReview ↗

More Table Results

Table A: Per-design WNS/TNS for the requested ablations

Format: WNS / TNS in ns; closer to zero is better.

DesignTP3Dno-LBP
ariane133−0.13 / −52.02−0.255 / −78.60
ariane136−1.07 / −2602.25−2.02 / −4737.18
bp_be−0.32 / −27.69−0.949 / −126.2
bp_fe−0.85 / −84.83−1.031 / −104.15
bp_multi−5.37 / −2837.15−6.264 / −4376.22
bp_quad−0.59 / −5250.15−0.68 / −7742.14
swerv_wrapper−0.39 / −134.75−0.642 / −261.9

Table B: Per-design Kendall τ

Kendall τ is computed within each design on that design's test placements. swerv_wrapper and bp_quad are held out of training.

DesignpWNS τpTNS τ
ariane1330.52230.6976
ariane1360.56210.6691
bp_be0.68070.5883
bp_fe0.53120.6529
bp_multi0.51040.6148
bp_quad (held-out)0.54750.6285
swerv_wrapper (held-out)0.59540.6437

Dear Reviewer LXtm,

We are the authors of the paper "TP3D: Timing-Aware Die-To-Die 3D Macro Placement". Thank you very much for the valuable comments you provided during the review. We would appreciate it if you could review our responses when you have a chance. We hope they address your concerns and clarify our work. If you find that your concerns have been sufficiently addressed, we would be grateful if you could consider raising your score. If any questions or concerns remain, please let us know. We would be very grateful if you could spare the time to reply.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.