Pith. sign in

REVIEW 4 major objections 4 minor 45 references

Recursive Learning-Based Virtual Buffering for Analytical Global Placement

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A learned recursive buffering model embedded in analytical global placement improves timing closure with no power penalty.

desk verdict Genuinely new open-source ML buffering integrated into analytical placement, but the no-power-degradation claim is contradicted by Table IV and the OpenROAD-flow evaluation is too thin. read the letter →

arxiv 2506.17247 v2 pith:U5J4S6JA submitted 2025-06-07 cs.LG cs.AI

Andrew B. Kahng , Yiting Liu , Zhiang Wang This is my paper · ORCID
classification cs.LGcs.AI
keywords virtualbufferinganalyticalglobalplacementgenerativemodelrecursivelearningbufferinsertionERCviolationstimingclosureOpenROAD
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper aims to show that a machine-learned buffer-placement model can be embedded directly inside analytical global placement and improve the quality of the finished chip. The authors build MLBuf, a recursive generative model that predicts where buffers should go and what size they should be for each electrical-rule-violating net, and integrate it into the RePlAce placer as MLBuf-RePlAce. The key move is virtual buffering: instead of physically inserting buffers during placement, MLBuf's predicted bounding boxes are converted into density occupancy, so the placer reserves space for buffers that later optimization stages will insert. Across open-source and commercial flows, the framework reports total negative slack (TNS) improvements of 56% at best and 31% on average in the OpenROAD flow, and 53% and 28% in a commercial flow, without degrading post-route power. If correct, this offers a cheaper, closed-loop alternative to repeatedly invoking classical buffering engines during placement.

What carries the argument

The load-bearing mechanism is MLBuf's recursive generative buffering model combined with the virtual-occupancy density update. MLBuf avoids Steiner tree construction through a differentiable clustering module built with Gumbel-Softmax soft assignments; two self-attention decoders then predict a buffer type and a location for each cluster, and the predictions are fed back as the next recursion level until no further buffers are needed. For placer integration, each predicted buffer bounding box $h_i$ overlapping a bin $b_j$ reduces that bin's allowed area by $A_{\text{overlap}}(h_i,b_j) \cdot D_j$, where $D_j$ is the bin's target density, so the density constraint pushes cells to spread out and leave room for future buffer insertion. The training signal combines inner-loop cluster, type, and location losses with outer-loop penalties for ERC violations, wirelength, and buffer area, allowing the model to search for solutions beyond its ground-truth labels.

What would settle it

Run the full OpenROAD flow on a design whose timing failures are dominated by maximum slew rather than by output capacitance (for example, long wires driven by strong buffers with small fanout), and compare MLBuf-RePlAce against TD-RePlAce; if slew violations remain high and the TNS improvement disappears, the slew-capacitance alignment assumption is the point of failure. A cheaper check is to retrain MLBuf with an explicit slew penalty and verify whether the residual slew violations reported in the per-iteration comparison drop.

Watch

Extended reading notes

Core claim

The paper's central claim is that recursive learning-based virtual buffering, coupled with bin-density virtual occupancy, can stand in for classical dynamic-programming buffering inside global placement. MLBuf treats each net as a sequence of drivers and sinks and iteratively predicts cluster assignments, buffer types, and buffer coordinates, using teacher forcing to stabilize training and constructing the full buffer-embedded tree auto-regressively at inference. The predicted buffer boxes feed a simple density bookkeeping rule that reduces the allowed area of each placement bin, which changes the density constraint that drives the placer. Compared with the default timing-driven placer in OpenROAD, the authors report no post-route power degradation and TNS improvements of 56% maximum and 31% average in the OpenROAD flow, and 53% maximum and 28% average in a commercial flow, with an average 0.2% post-route power improvement. They also state that this is the first ML-guided buffering approach assessed in the context of a complete placement-and-optimization flow.

Load-bearing premise

The model assumes that maximum slew violations track output capacitance closely enough that a capacitance-based ERC penalty suffices; if a design's slew violations grow independently of capacitance, the predicted buffer trees can leave electrical-rule violations unfixed and the claimed timing gains could vanish.

Editorial extensions

If this is right

  • If the central result holds, learned buffering can replace repeated calls to classical van Ginneken-Lillis-style engines during global placement, because MLBuf produces whole buffer-embedded trees without Steiner tree construction.
  • Because the density update reserves buffer space during placement, downstream resizing, buffering, and routing see less congestion and fewer detours, which is the paper's stated explanation for TNS gains at equal power.
  • MLBuf's inference is over 3x faster than OR rsz on large nets, so the cost of timing-driven guidance during placement drops as net sizes grow.
  • The held-out designs BlackParrot and MegaBoom show the learned model transfers to unseen designs, giving the framework a path to scalability beyond the designs used to collect training data.
  • Under very tight clock periods the benefit plateaus because MLBuf is ERC-driven and does not use timing slack, so its predicted buffer count becomes insufficient for extreme timing constraints.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable consequence the paper does not run: replacing MLBuf's predictions with OR rsz's actual buffer trees while applying the same bin-density update would isolate how much of the gain comes from prediction quality versus the porosity mechanism itself.
  • Because maximum slew is never explicitly penalized, the framework is likely to leave slew violations on nets where slew, not capacitance, is the binding electrical limit; adding an explicit slew penalty or slew-aware loss could close the residual slew-violation gap visible in the paper's Table III.
  • The per-net independence of MLBuf suggests a natural extension to GPU-accelerated placers: batching inference over all problematic nets could hide nearly all of the buffering cost inside the placer's parallel pipeline.
  • The paper's three-regime analysis (no benefit, consistent benefit, plateau) implies the model is most useful in a middle band of timing tightness; a dynamic trigger based on slack or overflow rather than a fixed overflow list could shift more designs into the beneficial regime.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes MLBuf, a recursive learning-based generative model that predicts buffer types and locations for nets with electrical rule check (ERC) violations, and integrates it into the RePlAce analytical global placer as MLBuf-RePlAce. Virtual buffering is achieved by pre-allocating bin occupancy according to predicted buffer bounding boxes, thereby making the placer buffer-porosity-aware. The framework is evaluated against the default timing-driven virtual-buffering placer in OpenROAD (TD-RePlAce) and an ad-hoc rule-based baseline, using both the open-source OpenROAD flow and a commercial Innovus flow, on five NanGate45 benchmarks. The central claims are that MLBuf-RePlAce improves post-route TNS by up to 56% (average 31%) in the OpenROAD flow without degrading post-route power, and by up to 53% (average 28%) in the commercial flow with a small average power improvement. The paper also contributes an open-source release of the code and training data.

Significance. If the claims hold, this is a valuable contribution to ML-guided physical design: it is, to my knowledge, the first learning-based virtual buffering method integrated into a full analytical placement and optimization loop, and the open-source release with OpenROAD integration is a strong practical asset. The recursive, teacher-forced training with differentiable clustering is a reasonable way to avoid expensive Steiner-tree construction. The authors also include held-out designs (BlackParrot and MegaBoom) in the commercial-flow evaluation, which partially addresses generalization concerns. However, the paper's headline claim—improved TNS 'without degradation of post-route power' in the OpenROAD flow—is directly contradicted by the paper's own reported numbers, and the OpenROAD-flow evaluation is restricted to the two training designs. These issues undermine the central claim as currently stated and require substantial revision.

major comments (4)
  1. [V-B, Table IV, Abstract] The abstract and Section V-B state that MLBuf-RePlAce improves TNS 'without degradation of post-route power' in the OpenROAD flow, but Table IV shows the opposite for jpeg: post-route power is 567.337 mW for MLBuf-RePlAce versus 559.573 mW for TD-RePlAce, a 1.4% increase. The average across the two reported OpenROAD-flow designs also increases from 343.70 mW to 347.34 mW. This is not a minor discrepancy: the 'without power degradation' qualifier is load-bearing for the paper's central claim. The claim should be revised to reflect a modest power tradeoff on jpeg, or the authors should provide error bars over multiple runs demonstrating that the difference is within run-to-run noise.
  2. [V and Appendix D2] The OpenROAD-flow PPA evaluation uses only ibex and jpeg, both of which contributed training data (Appendix D2 states that 70% of ibex and jpeg data is used for training). The paper excludes ariane because all of its data is used in training, and excludes BlackParrot and MegaBoom because OpenROAD fails to route them. Consequently, the headline open-flow TNS improvement is demonstrated only on training designs, and the baseline TD-RePlAce uses OR rsz, which is also the source of the ground-truth buffer-embedded trees for MLBuf. This creates a same-corpus evaluation risk. The authors should provide held-out OpenROAD-flow results (or clearly restrict the claim to training-seen designs) to support the generality implied by the abstract.
  3. [IV-C, Appendix B, Table III] The paper justifies not enforcing maximum slew constraints by asserting that slew effects are 'closely aligned with output capacitance' (Section IV-C and Appendix B). However, Table III shows that MLBuf leaves more slew violations than OR rsz on both ibex (58 vs 48) and jpeg (13 vs 5). Since slew violations are an explicit component of ERC violations, the empirical results contradict the stated rationale and weaken the claim that MLBuf 'addresses ERC violations' as effectively as the baseline. The authors should either incorporate a slew penalty or qualify the ERC claim by reporting slew violations as a known residual issue.
  4. [V-B] The denoising procedure averages PPA metrics over ±10 ps perturbations of the target clock period, but this does not provide a measure of variance across independent placement runs or optimization seeds. Given that the power differences in Table IV are small (e.g., ibex: 127.343 vs 127.828 mW), single-point averages are insufficient to establish 'no degradation.' The authors should report the spread of results (e.g., standard deviation or min/max over multiple runs) for the key PPA metrics, at least for the OpenROAD-flow designs.
minor comments (4)
  1. [IV-A] The sentence following Eq. (10) contains the typo 'correspoinding'; it should be 'corresponding.'
  2. [V-D, Figure 5] The caption of Figure 5 has formatting issues: 'clock f requency' and 'Type 2Consistent MLBuf benefits' should be 'clock frequency' and 'Type 2: Consistent MLBuf benefits.'
  3. [III, Eq. (2)] The variables hrx, hlx, hry, hly, brx, blx, bry, bly are used in Eq. (2) but are not explicitly defined in the text; a one-line definition would improve readability.
  4. [Table I] Table I lists a TCPInvs column with all entries '-', 'NA', or apparently masked values; the caption should state that TCPInvs is masked by design, since the body text refers to this masking only in a footnote.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: MLBuf's buffer-tree predictions are fit to OR rsz labels and then validated against OR rsz on the same designs; the OpenROAD-flow PPA headline is also reported only on training designs, although held-out commercial-flow results provide some independent grounding.

  1. fitted input called prediction [Section IV-B (Training Strategy) and Section V-A (Evaluation of Predicted Buffer-embedded Trees), Table III]
    "We use buffer-embedded trees generated by OpenROAD Resizer (OR rsz) during global placement ( repairDesign() in RepairDesign.cc [40]) as the ground truth. ... We compare the performance of buffer-embedded trees predicted by MLBuf and those calculated by OR rsz. Specifically, during each timing optimization, we insert the predicted buffers back into netlists according to the topology of the buffer-embedded trees on ibex and jpeg."

    MLBuf is trained with OR rsz's buffer-embedded trees as ground-truth labels, extracted from ibex, jpeg, and ariane. Section V-A then 'predicts' buffer-embedded trees and compares them against OR rsz on ibex and jpeg, i.e., the same designs and the same tool that produced the training labels. Table III's conclusion that MLBuf achieves comparable ERC violation resolution to OR rsz is therefore a goodness-of-fit to the training distribution, not an independent extrapolation. This is a supporting rather than central result: the full-flow TNS improvement is not determined solely by this fit because it depends on the density pre-allocation mechanism, so the circularity is partial.

full rationale

The central PPA claim does not reduce by definition to its inputs: MLBuf-RePlAce's virtual-occupancy density update (Eqs. 2-4) is an independent mechanism, and the commercial-flow results include fully held-out BlackParrot and MegaBoom designs, which give outside-the-training-corpus evidence. No uniqueness theorem is imported, and the self-citations (RePlAce, OpenROAD, DG-RePlAce) are to open-source tools and prior placer work, not to unverified results that carry the argument. The main circularity concern is the tree-quality evaluation in Table III, where the model's training labels (OR rsz outputs) are also the comparison baseline; 'comparable ERC resolution' there is fitted by construction. A second, non-circular validity concern is that the headline OpenROAD-flow improvement (56%/31% TNS) is reported only on ibex and jpeg, both of which contributed training data; ariane is excluded because all its OpenROAD data was used in training, and BP/MB are excluded because OpenROAD fails to route them. In addition, the abstract's 'without degradation of post-route power' is contradicted by Table IV for jpeg (567.337 mW vs 559.573 mW for TD-RePlAce), which is a correctness issue rather than a circularity issue. Overall, the derivation chain is not self-equivalent, but the fitted tree comparison and same-corpus evaluation justify a moderate circularity score.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a trained neural model with several undisclosed fitted parameters (alpha, beta, loss weights, Wmax), a clustering hyperparameter selected on the test set, and assumptions about ERC modeling and training labels. No new physical entities are introduced. These items are the main uncharged inputs.

free parameters (5)
  • Number of clusters k in the differentiable clustering module = 20
    Selected by sweeping k and evaluating ERC violations and buffer area on the test dataset (Appendix D1); tuning on test data can overstate performance.
  • Wirelength scaling factor alpha = not reported
    Appears in Eq. (11) for capacitance estimation and Eq. (15) for wirelength penalty; scales estimated wire resistance/capacitance, affecting buffer decisions, but its value and fitting method are not disclosed.
  • Output slew fitting factor beta = not reported
    Appears in Eq. (12); described as a fitting factor for higher-order delay effects, with no value given.
  • Wirelength penalty threshold Wmax = not reported
    Used in Eq. (15) to penalize long wires; the threshold value is not specified.
  • Loss and penalty weights (W_l1..W_l3, W_p1..W_p3, Warea) = not reported
    Hyperparameters in Algorithm 1 and Eq. (16) that balance the objective; values are not in the hyperparameter table.
assumptions (5)
  • domain assumption ERC violation detection via OpenSTA is accurate during global placement
    The flow relies on OpenSTA to identify nets with ERC violations (Section III); if timing or electrical estimates are inaccurate, MLBuf is trained and applied on wrong targets.
  • ad hoc to paper Maximum slew violations are implicitly covered by output capacitance penalties
    Appendix B penalty section states slew constraints are not explicitly enforced because their effects are 'closely aligned with output capacitance.' This is a modeling assumption not validated; Table III shows residual slew violations.
  • domain assumption Differentiable clustering can replace Steiner tree construction without degrading buffering quality
    Section IV-A avoids Steiner trees by soft clustering, but no comparison of tree quality against Steiner-based trees is shown.
  • domain assumption Ground-truth buffer-embedded trees from OR rsz are suitable training labels
    Section IV-B trains the model to imitate OpenROAD Resizer outputs, implicitly assuming those solutions are good references for end-to-end PPA.
  • domain assumption Self-attention over a net's cells captures the relevant spatial and electrical relationships
    Section IV-A uses self-attention encoders to model correlations among cells; this is a standard assumption behind the architecture.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Recursive Learning-Based Virtual Buffering for Analytical Global Placement." pith.science (2026). https://pith.science/paper/U5J4S6JA

@misc{pith2026250617247,
  author       = {Pith},
  title        = {Pith review of: Recursive Learning-Based Virtual Buffering for Analytical Global Placement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U5J4S6JA}},
  note         = {Machine review of arXiv:2506.17247}
}
read the original abstract

Due to the skewed scaling of interconnect versus cell delay in modern technology nodes, placement with buffer porosity (i.e., cell density) awareness is essential for timing closure in physical synthesis flows. However, existing approaches face two key challenges: (i) traditional van Ginneken-Lillis-style buffering approaches are computationally expensive during global placement; and (ii) machine learning-based approaches, such as BufFormer, lack a thorough consideration of Electrical Rule Check (ERC) violations and fail to "close the loop" back into the physical design flow. In this work, we propose MLBuf-RePlAce, the first open-source learning-driven virtual buffering-aware analytical global placement framework, built on top of the OpenROAD infrastructure. MLBuf-RePlAce adopts an efficient recursive learning-based generative buffering approach to predict buffer types and locations, addressing ERC violations during global placement. We compare MLBuf-RePlAce against the default virtual buffering-based timing-driven global placer in OpenROAD, using open-source testcases from the TILOS MacroPlacement and OpenROAD-flow-scripts repositories. Without degradation of post-route power, MLBuf-RePlAce achieves (maximum, average) improvements of (56%, 31%) in total negative slack (TNS) within the open-source OpenROAD flow. When evaluated by completion in a commercial flow, MLBuf-RePlAce achieves (maximum, average) improvements of (53%, 28%) in TNS with an average of 0.2% improvement in post-route power.

Figures

Figures reproduced from arXiv: 2506.17247 by the authors.

Figure 1
Figure 1. shows our approach. During timing-driven global place￾ment, when the placement overflow reaches a threshold specified in a predefined overflow list, a static timing engine (OpenSTA [35] in this work) is invoked to identify problematic nets with ERC violations [39]. Problematic nets, which also include long nets, are fed to the pre-trained MLBuf model for repair of ERC violations. (Details of MLBuf are given in Secti… view at source ↗
Figure 2
Figure 2. MLBuf’s bottom-up learning process. embedded tree with low accuracy. To mitigate this issue, we adopt the teacher forcing strategy [29] during training, which helps stabilize the learning process and reduce error propagation across recursive iterations (see Section IV-B). Third, to encourage MLBuf to generate superior buffering so￾lutions rather than just replicating ground-truth, we propose a specialized training p… view at source ↗
Figure 3
Figure 3. Model structure of MLBuf. B. Training Strategy This section describes the training and inference strategy for ML￾Buf. We use buffer-embedded trees generated by OpenROAD Re￾sizer (OR rsz) during global placement (repairDesign() in RepairD￾esign.cc [40]) as the ground truth. We define the level of a cell in the buffer-embedded tree as the longest path from itself to the sink. Suppose a buffer-embedded tree has H˜ leve… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Runtime comparison between MLBuf and OR rsz [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Post-route TNS comparisons between MLBuf-RePlAce and TD-RePlAce under different clock frequencies [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Hyperparameter evaluation. The normalized number of ERC [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 42 canonical work pages

  1. [1]

    GOALPlace: Begin with the End in Mind

    A. Agnesina, R. Liang, G. Pradipta, A. Rajaram and H. Ren, “GOALPlace: Begin with the End in Mind”, Proc. ISPD , 2025, pp. 2-10

    work page 2025

  2. [2]

    Toward an Open-Source Digital Flow: First Learnings from the OpenROAD Project

    T. Ajayi, V . A. Chhabria, M. Fogac ¸a, S. Hashemi, A. Hosny, A. B. Kahng, M. Kim, J. Lee, U. Mallappa, M. Neseem, G. Pradipta, S. Reda, M. Saligane, S. S. Sapatnekar, C. Sechen, M. Shalan, W. Swartz, L. Wang, Z. Wang, M. Woo and B. Xu, “Toward an Open-Source Digital Flow: First Learnings from the OpenROAD Project”, Proc. DAC, 2019, pp. 1-4

    work page 2019

  3. [3]

    Porosity-Aware Buffered Steiner Tree Construction

    C. J. Alpert, G. Gandham, M. Hrkic, Jiang Hu, S. T. Quay and C. N. Sze, “Porosity-Aware Buffered Steiner Tree Construction”, IEEE Trans. on CAD 23(4) (2004), pp. 517-526

    work page 2004

  4. [4]

    Buffered Steiner Trees for Difficult Instances

    C. J. Alpert, M. Hrkic, J. Hu, A. B. Kahng, J. Lillis, B. Liu, S. T. Quay, S. S. Sapatnekar and A. J. Sullivan, “Buffered Steiner Trees for Difficult Instances”, Proc. ISPD, 2021, pp. 4-9

    work page 2021

  5. [5]

    Techniques for Fast Physical Synthesis

    C. J. Alpert, S. K. Karandikar, Z. Li, G.-J. Nam, S. T. Quay and H. Ren, “Techniques for Fast Physical Synthesis”, Proc. of the IEEE 95(3) (2007), pp. 573-599

    work page 2007

  6. [6]

    On Wirelength Estimations for Row-Based Placement

    A. E. Caldwell, A. B. Kahng, S. Mantik, I. L. Markov and A. Zelikovsky, “On Wirelength Estimations for Row-Based Placement”, IEEE Trans. on CAD 18(9) (1999), pp. 1265-1278

    work page 1999

  7. [7]

    An Integrated Nonlinear Placement Framework with Congestion and Porosity Aware Buffer Planning

    T.-C. Chen, A. Chakraborty and D. Z. Pan, “An Integrated Nonlinear Placement Framework with Congestion and Porosity Aware Buffer Planning”, Proc. DAC, 2008, pp. 702-707

    work page 2008

  8. [8]

    RePlAce: Advanc- ing Solution Quality and Routability Validation in Global Placement

    C.-K. Cheng, A. B. Kahng, I. Kang and L. Wang, “RePlAce: Advanc- ing Solution Quality and Routability Validation in Global Placement”, IEEE Trans. on CAD 38(9) (2019), pp. 1717-1730

    work page 2019

Show all 45 references
  1. [9]

    FLUTE: Fast Lookup Table Based Rectilin- ear Steiner Minimal Tree Algorithm for VLSI Design

    C. Chu and Y . -C. Wong, “FLUTE: Fast Lookup Table Based Rectilin- ear Steiner Minimal Tree Algorithm for VLSI Design”, IEEE Trans. on CAD 27(1) (2008), pp. 70-83

    2008

  2. [10]

    Routing Tree Construction Under Fixed Buffer Locations

    J. Cong and X. Yuan, “Routing Tree Construction Under Fixed Buffer Locations”, Proc. DAC, 2000, pp. 379–384

    2000

  3. [11]

    Fusion of Global Placement and Gate Sizing with Differentiable Optimization

    Y . Du, Z. Guo, Y . Lin, R. Wang and R. Huang, “Fusion of Global Placement and Gate Sizing with Differentiable Optimization”, Proc. ICCAD, 2024, pp. 1-9

    2024

  4. [12]

    The Transient Response of Damped Linear Networks with Particular Regard to Wideband Amplifiers

    W. C. Elmore, “The Transient Response of Damped Linear Networks with Particular Regard to Wideband Amplifiers”, Journal of Applied Physics 19(1) (1948), pp. 55-63

    1948

  5. [13]

    Differentiable-Timing-Driven Global Placement

    Z. Guo and Y . Lin, “Differentiable-Timing-Driven Global Placement”, Proc. DAC, 2022, pp. 1315 - 1320

    2022

  6. [14]

    Timing-Driven Placement Based on Dynamic Net-Weighting for Efficient Slack Histogram Compression

    C. Guth, V . Livramento, R. Netto, R. Fonseca, J. L. G ¨untzel and L. Santos, “Timing-Driven Placement Based on Dynamic Net-Weighting for Efficient Slack Histogram Compression”, Proc. ISPD , 2015, pp. 141–148

    2015

  7. [15]

    Gaussian Error Linear Units (GELUs)

    D. Hendrycks and K. Gimpel, “Gaussian Error Linear Units (GELUs)”, arXiv preprint arXiv:1606.08415 , 2016

    2016 arXiv

  8. [16]

    Categorical Reparameterization with Gumbel-Softmax

    E. Jang, S. Gu and B. Poole, “Categorical Reparameterization with Gumbel-Softmax”, Proc. ICLR, 2017, pp. 1-12

    2017

  9. [17]

    DG-RePlAce: A Dataflow-Driven GPU- Accelerated Analytical Global Placement Framework for Machine Learning Accelerators

    A. B. Kahng and Z. Wang, “DG-RePlAce: A Dataflow-Driven GPU- Accelerated Analytical Global Placement Framework for Machine Learning Accelerators”, IEEE Trans. on CAD 44(2) (2025), pp. 696- 708

    2025

  10. [18]

    BufFormer: A Generative ML Framework for Scalable Buffering

    R. Liang, S. Nath, A. Rajaram, J. Hu, and H. Ren, “BufFormer: A Generative ML Framework for Scalable Buffering”, Proc. ASP-DAC, 2023, pp. 264-270

    2023

  11. [19]

    DREAMPlace 4.0: Timing-Driven Global Placement with Momentum-Based Net Weighting

    P. Liao, S. Liu, Z. Chen, W. Lv, Y . Lin and B. Yu, “DREAMPlace 4.0: Timing-Driven Global Placement with Momentum-Based Net Weighting”, Proc. DATE, 2022, pp. 939–944

    2022

  12. [20]

    Simultaneous Routing and Buffer Insertion for High Performance Interconnect

    J. Lillis, C.-K. Cheng and T.-T. Y . Lin, “Simultaneous Routing and Buffer Insertion for High Performance Interconnect”, Proc. GLSVLSI, 1996, pp. 148-153

    1996

  13. [21]

    Focal Loss for Dense Object Detection

    T.-Y . Lin, P. Goyal, R. Girshick, K. He and P. Dollar, “Focal Loss for Dense Object Detection”, IEEE Trans Pattern Anal Mach Intell 42(2), 2020, pp. 318-327

    2020

  14. [22]

    Electrostatics-Based Analytical Global Placement for Timing Opti- mization

    Z. Lin, M. Wei, Y . Chen, P. Zou, J. Chen and Y .-W. Chang, “Electrostatics-Based Analytical Global Placement for Timing Opti- mization”, Proc. DATE, 2024, pp. 1-6

    2024

  15. [23]

    ePlace: Electrostatics-Based Placement Using Fast Fourier Transform and Nesterov’s Method

    J. Lu, P. Chen, C.-C. Chang, S. Lu, D. J.-H. Huang, C.-C Teng and C.- K. Cheng, “ePlace: Electrostatics-Based Placement Using Fast Fourier Transform and Nesterov’s Method”, ACM Trans. Des. Autom. Electron. Syst. 20(2) (2015), pp. 17:1-17:34

    2015

  16. [24]

    A Novel Technique Integrating Buffer Insertion into Timing Driven Placement

    L. Luo, Q. Zhou, Y . Cai, X. Hong and Y . Wang, “A Novel Technique Integrating Buffer Insertion into Timing Driven Placement”, Proc. ISCAS, 2006, pp. 1-4

    2006

  17. [25]

    RUMBLE: An Incremental Timing-Driven Physical-Synthesis Optimization Algorithm

    D. A. Papa, T. Luo, M. D. Moffitt, C. N. Sze, Z. Li and G.-J. Nam, “RUMBLE: An Incremental Timing-Driven Physical-Synthesis Optimization Algorithm”, IEEE Trans. on CAD27(12), 2008, pp. 2156- 2168

    2008

  18. [26]

    Autonomous Application of Netlist Transformations Inside Lagrangian Relaxation-Based Optimization

    A. Stefanidis, D. Mangiras, C. Nicopoulos, D. Chinnery and G. Dimitrakopoulos, “Autonomous Application of Netlist Transformations Inside Lagrangian Relaxation-Based Optimization”, IEEE Trans. on CAD 40(8) (2021), pp. 1672-1686

    2021

  19. [27]

    Buffer Placement in Distributed RC-tree Networks for Minimal Elmore Delay

    L. P. P. P. Van Ginneken, “Buffer Placement in Distributed RC-tree Networks for Minimal Elmore Delay”, Proc. ISCAS , 1990, pp. 865- 868

    1990

  20. [28]

    Attention Is All You Need

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser and I. Polosukhin, “Attention Is All You Need”, Proc. NeurIPS, 2017, pp. 6000-6010

    2017

  21. [29]

    A Learning Algorithm for Continually Running Fully Recurrent Neural Networks

    R. J. Williams and D. Zipser, “A Learning Algorithm for Continually Running Fully Recurrent Neural Networks”, Neural computation 1(2) (1989), pp. 270–280

    1989

  22. [30]

    AiTO: Simultaneous gate sizing and buffer insertion for timing optimization with GNNs and RL

    H. Wu, Z. Huang, X. Li and W. Zhu, “AiTO: Simultaneous gate sizing and buffer insertion for timing optimization with GNNs and RL”, Integration 98 (2024), pp. 102211

    2024

  23. [31]

    Hierarchical Graph Representation Learning with Differentiable Pool- ing

    R. Ying, J. You, C. Morris, X. Ren, W. L. Hamilton and J. Leskovec, “Hierarchical Graph Representation Learning with Differentiable Pool- ing”, Proc. NeurIPS, 2018, pp. 4805–4815

    2018

  24. [32]

    https://github.com/The-OpenROAD-Project/OpenROAD/tree/master/ src/rsz

    Gate Resizer. https://github.com/The-OpenROAD-Project/OpenROAD/tree/master/ src/rsz

  25. [33]

    https://github.com/The-OpenROAD-Project/OpenROAD-flow-scripts

    OpenROAD-Flow-Scripts, commit hash: 52d002 . https://github.com/The-OpenROAD-Project/OpenROAD-flow-scripts

  26. [34]

    https://github.com/The-OpenROAD-Project/OpenROAD

    OpenROAD, commit hash: df581be . https://github.com/The-OpenROAD-Project/OpenROAD

  27. [35]

    https://github.com/The-OpenROAD-Project/OpenSTA/tree/master

    OpenSTA. https://github.com/The-OpenROAD-Project/OpenSTA/tree/master

  28. [36]

    https://github.com/TILOS-AI-Institute/MacroPlacement

    TILOS MacroPlacement repository. https://github.com/TILOS-AI-Institute/MacroPlacement

  29. [37]

    https://eda.ncsu.edu/freepdk/freepdk45/

    NanGate45 PDK. https://eda.ncsu.edu/freepdk/freepdk45/

  30. [38]

    https://github.com/ABKGroup/MLBuf MLCAD

    The MLBuf repository. https://github.com/ABKGroup/MLBuf MLCAD

  31. [39]

    https://github.com/The-OpenROAD-Project/OpenROAD/tree/master/ src/gpl

    The global placement module in OpenROAD (gpl). https://github.com/The-OpenROAD-Project/OpenROAD/tree/master/ src/gpl

  32. [40]

    https://github.com/The-OpenROAD-Project/OpenROAD/tree/master/ src/rsz

    The gate resizer module in openROAD (rsz). https://github.com/The-OpenROAD-Project/OpenROAD/tree/master/ src/rsz. APPENDIX A. Terminology and Notation Table VI summarizes all terms and their definitions. TABLE VI: Terminology and Notation. Basic Entities v Cell (driver, sink o...

  33. [41]

    Algorithm 1 MLBuf Training Procedure Input: Training set Dtrain = {T1,

    and model parameters are updated via back-propagation (Line 16). Algorithm 1 MLBuf Training Procedure Input: Training set Dtrain = {T1, . . . , TN }, maximum tree depth ˜H, learnable parameters δ, learning rate η Output: Optimized model parameters δ⋆ 1: for each tree T ∈ Dtrai...

  34. [42]

    Hyperparameter Selection: The most important hyperparam- eter in MLBuf is the number of clusters k (i.e., the output dimension of the clustering module). It affects the structure of the predicted Algorithm 2 MLBuf Inference Procedure Input: Driver d and initial sink set V s 0,...

  35. [43]

    We use 70% of the collected data from ibex and jpeg, along with all the data from ariane, to train the model

    MLBuf Training Details: The training dataset is collected from OpenROAD using the data collection process described in Section V. We use 70% of the collected data from ibex and jpeg, along with all the data from ariane, to train the model. 20% of the ibex and jpeg data is used...

  36. [44]

    Specifically, we insert the predicted buffers back into the netlist based on the predicted buffer-embedded tree topology in each timing optimization iteration (13 times)

    Performance Comparison of Buffer-embedded Trees: We com- pare the performance of buffer-embedded trees predicted by MLBuf and those calculated by OR rsz. Specifically, we insert the predicted buffers back into the netlist based on the predicted buffer-embedded tree topology in...

  37. [45]

    non-uniformly

    Ad-Hoc Baseline: This is a rule-based analytical approxima- tion used as a baseline for comparison, as discussed in Section V-B. The approach involves (i) estimating the net wirelength using a wireload model (WLM) [6]; (ii) applying a pessimism margin of 25% to the estimated w...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.