LEAP: A Self-Supervised Per-Cycle Toggle Propagation Model Supports Fast, Transferable, and Early Analysis of Layout Power

Mengming Li; Wenji Fang; Wenkai Li; Yao Lu; Yuchao Wu; Zhiyao Xie; Ziyan Guo

REVIEW 3 major objections 5 minor 26 references

LEAP: A Self-Supervised Per-Cycle Toggle Propagation Model Supports Fast, Transferable, and Early Analysis of Layout Power

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read LEAP sets out to prove that per-cycle toggle propagation through combinational logic can be learned by a linear-complexity graph transformer, replacing the slowest step in layout power analysis.

desk verdict First serious per-cycle toggle propagation model, but unspecified pre-training split could invalidate the cross-design claim. read the letter →

arxiv 2608.01946 v1 pith:HK74DG52 submitted 2026-08-03 cs.AR

Wenkai Li , Yuchao Wu , Ziyan Guo , Yao Lu , Wenji Fang , Mengming Li , Zhiyao Xie This is my paper

classification cs.AR

keywords per-cycletogglepropagationgraphtransformerlinear-causalattentionself-supervisedpre-traininglayoutpowerestimationnetlistanalysisVLSItransferlearning

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

LEAP claims that per-cycle toggle propagation—the step that extends register activity through combinational logic in a netlist—can be replaced by a learned model without losing accuracy. The authors build a graph transformer whose attention is causal in topological depth, so each gate only sees signals from earlier logic stages, mirroring how toggles actually travel. With self-supervised pre-training on gate function and structure, the model reaches PR-AUC 0.99 for predicting which combinational gates toggle each cycle, 7.6x faster than the commercial EDA tool. Integrated with an existing ML power model, it yields per-cycle layout power from post-synthesis netlists at 4.55% MAPE and 5.3x lower total runtime. If these numbers hold, per-cycle power analysis no longer needs the layout and toggle-propagation bottleneck.

What carries the argument

The load-bearing mechanism is LEAP-Former's linear-causal attention: each node's query attends only to keys and values accumulated from all nodes at strictly shallower topological depths, with cumulative sums S and Z updated depth-by-depth, so complexity stays linear rather than quadratic. A parallel Gated GCN branch applies learned gates to neighbor messages, so not all input signals are treated equally, and 128-dimensional text embeddings of each gate's Boolean expression give the model functional knowledge. These pieces let the model propagate register toggle status through combinational logic in one forward pass matching signal direction.

What would settle it

Run a single experiment: train on designs D1–D8 and test exclusively on designs D9–D12 (design IDs never seen during training) across all three workloads, and report per-cycle PR-AUC. If PR-AUC drops below 0.9 while the same model still hits 0.99 on seen-design workloads, the cross-design claim is falsified and the true contribution is cross-workload generalization within a design family.

Watch

Extended reading notes

Core claim

On the paper's own terms, LEAP is the first per-cycle toggle propagation predictor. The model converts a post-synthesis netlist into a directed graph and uses a single-layer, multi-head attention where each node attends only to nodes at lower topological depths; cumulative sums keep the cost linear in graph size. Three pre-training tasks—predicting per-cycle combinational output values, regressing static switching probabilities, and reconstructing masked functional embeddings—teach the encoder circuit behavior before supervised fine-tuning on toggle labels. On twelve out-of-order CPUs with 240k–1.2M gates, LEAP attains near-perfect PR-AUC (0.99) and ROC-AUC (1.00) on three workloads, runs 7.

Load-bearing premise

The load-bearing premise is that toggle behavior in a gate-level netlist is fully determined by local gate function, topology, and register activity, and that a model trained on 12 CPUs with design-overlapping validation generalizes to unseen designs; the 2-fold validation does not hold out whole design IDs, so this transfer is assumed, not yet demonstrated.

Editorial extensions

If this is right

If LEAP works as reported, designers can obtain per-cycle toggle information from post-synthesis netlists without running the full RTL-to-layout power simulation, cutting the most expensive step of power analysis.
The 7.6x toggle-propagation speedup and 5.3x end-to-end power-analysis speedup would allow many more power checks per day, making per-cycle power data practical for design-space exploration.
Because inputs are post-synthesis netlists plus RTL simulation, power estimates become available before physical design, enabling early power-driven optimization.
The same learned toggle propagation could feed dynamic timing analysis and glitch-power estimation, not just average-power models.
The linear attention plus gated-GCN structure suggests the approach can scale to larger netlists beyond the 1.2M-gate CPUs tested.

Reading between the lines

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

The reported 2-fold validation trains and tests on the same twelve design IDs—only the workload changes—so the near-perfect PR-AUC does not yet demonstrate transfer to completely new netlist architectures; a holdout of entire design IDs would tighten the claim.
If the functional embeddings capture gate behavior, the pre-training might transfer across standard-cell libraries with only light fine-tuning, since the Boolean expressions are library-agnostic text.
The self-supervised tasks are generic enough that the same LEAP-Former backbone could be repurposed for other per-cycle netlist phenomena, such as glitch rate or critical-path timing, where labels are equally expensive to obtain.
A testable extension: train on a subset of designs and report PR-AUC on completely unseen designs; if it stays near 0.99, the cross-design story is confirmed; if not, the real regime is cross-workload only.

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.

Desk Editor's Note

read the letter

Bottom line: this is the first paper I've seen that seriously targets per-cycle toggle propagation, and the architecture is a sensible adaptation of linear attention to the depth direction of a netlist. But the evaluation has a hole the authors haven't acknowledged. The pre-training section (5.1) never says which designs are used for the self-supervised tasks, and the 2-fold validation in Section 6.2 only defines the split for fine-tuning. If pre-training runs on all 12 designs, then for each fold the model has already seen the test designs' netlists and their per-cycle output pin values during task L#1. When you have a frozen Value_MLP that predicts those values and is then fed into the toggle classifier, near-perfect PR-AUC could simply mean the model memorized the test designs, not that it learned toggle propagation. This doesn't have to be a fatal problem, but the paper doesn't rule it out.

What's genuinely good: the task itself is worth doing; prior work (Grannite, DeepSeq) only predicted average toggle rates. The depth-wise linear-causal attention is a clean way to respect directionality, and the paper is honest that it's a single attention layer. Pre-training with value prediction, static switching regression, and masked functional embedding is standard but applied here to a new domain. The evaluation on 12 out-of-order CPUs is more realistic than most ML-EDA papers. The ablations are informative: removing register toggle features collapses PR-AUC to 0.12, which shows the model is actually using the propagation source.

Softer spots: no variance over seeds, which makes the 0.99 point estimates hard to judge; the Grannite baseline is not a real comparison since it wasn't designed for per-cycle toggles; only 12 designs, all out-of-order CPUs, so generalization to other microarchitectures is untested; power results are only for Setting 1.

Verdict: the paper deserves a serious referee. The central claim of transferable per-cycle toggle propagation is plausible but not established. I'd want the authors to state the pre-training data split, rerun with pre-training restricted to the training fold if needed, and report seed variance. If that confirms the numbers, this is an accept; if not, the cross-design claim falls.

Recommendation: send it out, with a request to address the split explicitly.

Referee Report

3 major / 5 minor

Summary. The paper proposes LEAP, a per-cycle toggle propagation prediction model for gate-level netlists. LEAP uses a linear-complexity graph transformer with a novel linear-causal attention mechanism that restricts each node to information from earlier topological depths, mimicking signal propagation. A self-supervised pre-training stage (tasks for output-pin-value prediction, static switching-characteristic regression, and functional embedding reconstruction) is followed by supervised fine-tuning to predict per-node toggles. The authors report near-perfect PR-AUC (0.99) across 12 out-of-order CPU designs and 3 workloads, a 7.6x runtime speedup over the Synopsys PTPX tool, and integration with the ATLAS power model (LEAP-Power) to achieve 4.55% MAPE for per-cycle layout power. The paper claims this is the first approach to accurately predict per-cycle toggle propagation and to enable fast, transferable, early layout power analysis.

Significance. If the central claims hold, the contribution is significant: a linear-complexity graph transformer with a causal attention mechanism designed for directed netlists, combined with self-supervised pre-training, would address a real bottleneck in per-cycle power analysis. The evaluation uses realistic large-scale out-of-order CPU designs and compares against a commercial EDA tool, which strengthens practical relevance. The integration with ATLAS demonstrates a concrete downstream use case. However, the cross-design generalization claim, which is a central selling point, depends critically on the pre-training data split, and the paper currently does not specify it. The reported near-perfect PR-AUC values also lack any measure of variance, making it difficult to assess robustness. The architecture and pre-training ideas are novel and worth pursuing, but the current experimental validation is incomplete.

major comments (3)

[§5.1 and §6.2] The pre-training data split is not specified. Section 5.1 describes the pre-training tasks but does not state which designs or cycles are used. Section 6.2 defines a 2-fold fine-tuning split, but if pre-training uses all 12 designs (including those later assigned to the test fold), then for each fold the test designs' netlists, per-cycle output pin values (Task #1), and static switching characteristics have already been observed during pre-training. Since the Value_MLP is frozen and its predictions are used as an input feature during fine-tuning (§5.2), the model can effectively access test-design behavior, undermining the cross-design generalization claim. Please specify the pre-training split; if it includes test designs, re-run the experiments with pre-training restricted to the fine-tuning training fold and report results for fully held-out designs.
[Table 2, §6.2] All PR-AUC and ROC-AUC values are reported as single numbers (e.g., 0.99, 1.00) with no variance over random seeds or multiple runs. Given the class imbalance (3-10% positive samples) and the small number of designs (12), the absence of error bars makes it impossible to assess statistical significance or the stability of the claimed 0.99 PR-AUC. Please report mean and standard deviation over at least five seeds, and state the number of seeds used.
[§6.5, Figure 4] The ablation that removes toggle-related features and observes a PR-AUC drop to 0.12 is expected and confirms the model relies on register toggles as initial conditions. However, this ablation does not address the core generalization question: whether the model can predict toggles for designs whose structure and behavior were never seen in either pre-training or fine-tuning. A proper cross-design evaluation must exclude entire designs from both stages; the current 2-fold scheme only excludes from fine-tuning but potentially not from pre-training.

minor comments (5)

[Title and abstract] The title contains a typo: 'Sel f-Supervised' should be 'Self-Supervised'. The abstract is otherwise clear.
[§6.3, Table 3] The sentence 'while its acceleration over Grannite (9.3×) is slightly lower' is misleading. Grannite is 9.3× faster than PTPX, not 'acceleration over Grannite'. Please rephrase.
[§4.2, Eq. (2)-(4)] The notation in Eq. (2) is dense. It would help to define the dimensions of Q, K, V and the summation indices explicitly, and to state that the sums in Eq. (3) run over all nodes at all depths before d, not just depth d-1.
[§6.1] The paper does not mention the total number of cycles per workload beyond '300-cycle segments'. Is this the full workload length or a subsample? Please clarify whether the 300 cycles are consecutive and whether multiple segments per design were used.
[§5.1] The description of the masked functional embedding reconstruction (Task #3) does not specify the masking ratio. Please state it, along with all pre-training hyperparameters (e.g., alpha coefficients, learning rate, batch size).

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: LEAP's toggle predictions are learned from explicit per-cycle supervision; the only concern is an unspecified pretraining data split, which is a generalization risk, not a by-construction circularity.

full rationale

LEAP's claimed derivation is an empirical supervised-learning pipeline, not a closed-form derivation. Eq. (1) defines toggle propagation as the mapping f_toggle(P_n, Tog_reg,t) -> Tog_comb,t, and the reported PR-AUC is measured against ground-truth PTPX toggle labels. The linear-causal attention (Eqs. 2–4) is an architectural inductive bias that restricts information flow by topological depth; it mimics toggle propagation but does not redefine the target, and its linear-attention core is cited to external prior work [12], with GCN/gating from external sources [4,15,21]. The Value_MLP from Task L#1 is an auxiliary multi-task head: it is frozen and its output is concatenated as an extra feature for the toggle classifier, which is trained with supervised toggle labels. The ablation study shows that removing the entire pre-training stage (and thus Value_MLP) leaves PR-AUC at 0.93, so the final prediction does not reduce to the auxiliary value prediction by construction. Self-citations to ATLAS and NetTAG are used for downstream integration and feature extraction, not to justify the core toggle-propagation result; no 'uniqueness theorem' from the authors is invoked. The one genuine concern is an omitted specification: Section 5.1 says pretraining uses 'the workload' to obtain per-cycle output-pin values but does not state whether the 12 designs are split so that the fine-tuning test fold is excluded from pretraining. If pretraining included test cycles, the frozen Value_MLP could leak value-level information and inflate PR-AUC, but this is a data-leakage/generalization risk, not a by-construction circularity; it is checkable and fixable. Accordingly, no circular step is exhibited.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The model uses standard deep learning components and no new physical or conceptual entities. The only ad-hoc elements are the evaluation split and the manually chosen pre-training loss weights.

free parameters (2)

Pre-training loss coefficients alpha_value, alpha_p, alpha_recon
Manually chosen weights in Eq. 5; values not reported.
Model hyperparameters (attention heads, hidden dims, MLP layers)
Not reported; affect performance and runtime.

assumptions (4)

standard math The linear attention formulation from Katharopoulos et al. (Eq. 2) is valid for graph nodes grouped by topological depth.
Used in Section 4.2 to derive depth-wise attention.
domain assumption Netlists are acyclic so topological depth is well-defined and every combinational node has finite depth.
Section 3 defines topological depth as max distance to input registers; assumes no combinational loops.
domain assumption Toggle behavior at a gate is determined by the register toggles and the gate's Boolean function, so depth-wise causal attention is a sufficient inductive bias.
Central modeling assumption in Section 4.2.
ad hoc to paper The 2-fold validation split with shared designs across train/test is a valid measure of cross-design generalization.
Section 6.2 Settings 1 and 2 include the same designs in both training and testing with different workloads.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LEAP: A Self-Supervised Per-Cycle Toggle Propagation Model Supports Fast, Transferable, and Early Analysis of Layout Power." pith.science (2026). https://pith.science/paper/HK74DG52

@misc{pith2026260801946,
  author       = {Pith},
  title        = {Pith review of: LEAP: A Self-Supervised Per-Cycle Toggle Propagation Model Supports Fast, Transferable, and Early Analysis of Layout Power},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HK74DG52}},
  note         = {Machine review of arXiv:2608.01946}
}

read the original abstract

Accurate power analysis is critical in VLSI design, as it directly impacts power optimization strategies. However, traditional approaches are often hindered by the substantial runtime required for per-cycle toggle propagation in the netlist, which propagates register toggle information through combinational logic. To address this, we propose LEAP, the first work to enable per-cycle toggle propagation prediction with both high accuracy and efficiency. This is achieved through a novel, linear-complexity graph transformer capable of simulating toggle propagation, along with specially designed self-supervised pre-training tasks that enable the model to capture circuit structure and functionality. LEAP achieves a 7.6x speedup over the EDA tool in toggle propagation, and attains a near-perfect area under the Precision-Recall curve (PR-AUC) of 0.99 for prediction results. Moreover, LEAP can be seamlessly integrated with other machine learning based power models into LEAP-Power. This integration enables precise per-cycle layout power prediction directly from post-synthesis netlists, achieving a mean absolute percentage error(MAPE) of only 4.55%. By bypassing toggle propagation in the netlist, LEAP-Power delivers substantial runtime gains, running 5.3x faster than the model without LEAP.

Figures

Figures reproduced from arXiv: 2608.01946 by the authors.

**Figure 2.** LEAP Overview. LEAP includes three major steps: dataset processing (Section 3), pre-training (Section 5.1), and [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** LEAP-Former Overview. The core of LEAP-Former is a linear-causal attention mechanism designed for directed graphs to mimic toggle propagation. It operates in two main steps: 𝑆 and 𝑍 updating, followed by attention calculation (Section 4.2). In addition, to convey structural information and thereby avoid the need for positional embeddings, a Gated GCN is integrated into the model (Section 4.3). The combined node embe… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Ablation study on pre-training and feature types. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 1 linked inside Pith

[1]

https://www.tsmc.com/english/dedic atedFoundry/technology/logic/l_40nm

2008.TSMC 40nm LP process technology. https://www.tsmc.com/english/dedic atedFoundry/technology/logic/l_40nm

2008
[2]

Design Compiler®RTL Synthesis

2021. Design Compiler®RTL Synthesis. https://www.synopsys.com/ implementation-and-signoff/rtl-synthesis-test/design-compiler-nxt.html

2021
[3]

VCS®functional verification solution

2021. VCS®functional verification solution. https://www.synopsys.com/ verification/simulation/vcs.html

2021
[4]

Xavier Bresson and Thomas Laurent. 2018. Residual Gated Graph ConvNets

2018
[5]

Cadence. 2021. Innovus Implementation System

2021
[6]

Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu
[7]

Djork-Arné Clevert, Thomas Unterthiner, and Sepp Hochreiter. 2015. Fast and ac- curate deep network learning by exponential linear units (elus).arXiv:1511.07289 (2015)

arXiv 2015
[8]

Jesse Davis and Mark Goadrich. 2006. The relationship between Precision-Recall and ROC curves. InProceedings of the 23rd International Conference on Machine Learning

2006

Show all 26 references

[9]

Yufan Du et al. 2024. PowPrediCT: Cross-Stage Power Prediction with Circuit- Transformation-Aware Learning. InProc. Design Automation Conf. (DAC). 1–6

2024
[10]

Wenji Fang et al. 2023. MasterRTL: A Pre-Synthesis PPA Estimation Framework for Any RTL Design. InProc. IEEE/ACM Int. Conf. Comput. Aided Design (ICCAD),. 1–9

2023
[11]

Wenji Fang, Wenkai Li, Shang Liu, Yao Lu, Hongce Zhang, and Zhiyao Xie
[12]

Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret
[13]

Sadaf Khan, Zhengyuan Shi, Min Li, and Qiang Xu. 2024. DeepSeq: Deep Sequen- tial Circuit Learning. In2024 Design, Automation & Test in Europe Conference & Exhibition (DATE)

2024
[14]

Wenkai Li, Yao Lu, Wenji Fang, Jing Wang, Qijun Zhang, and Zhiyao Xie. 2025. ATLAS: A Self-Supervised and Cross-Stage Netlist Power Model for Fine-Grained Time-Based Layout Power Analysis. InProceedings of the 62th ACM/IEEE Design Automation Conference

2025
[15]

Yujia Li, Richard Zemel, Marc Brockschmidt, and Daniel Tarlow. 2016. Gated Graph Sequence Neural Networks. InProceedings of ICLR’16

2016
[16]

Rongjian Liang, Zhiyao Xie, Jinwook Jung, Vishnavi Chauha, Yiran Chen, Jiang Hu, Hua Xiang, and Gi-Joon Nam. 2020. Routing-free crosstalk prediction. In International Conference on Computer Aided Design (ICCAD)

2020
[17]

Huizi Mao, Xiaodong Yang, and William J Dally. 2019. A delay metric for video ob- ject detection: What average precision fails to tell. InProceedings of the IEEE/CVF International Conference on Computer Vision

2019
[18]

Synopsys. 2023. PrimePower: RTL to Signoff Power Analysis. https://www. synopsys.com/implementation-and-signoff/signoff/primepower.html

2023
[19]

Gomez, Łukasz Kaiser, and Illia Polosukhin

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. InProceedings of the 31st International Conference on Neural Information Processing Systems

2017
[20]

Qitian Wu et al. 2023. DIFFormer: Scalable (Graph) Transformers Induced by En- ergy Constrained Diffusion. InThe Eleventh International Conference on Learning Representations

2023
[21]

Qitian Wu et al. 2023. SGFormer: Simplifying and Empowering Transformers for Large-Graph Representations. InAdvances in Neural Information Processing Systems (NeurIPS)

2023
[22]

Qitian Wu, Wentao Zhao, et al. 2024. NodeFormer: a scalable graph structure learning transformer for node classification. InProceedings of the 36th Interna- tional Conference on Neural Information Processing Systems

2024
[23]

Yanqing Zhang et al. 2020. GRANNITE: Graph neural network inference for transferable power estimation. InProc. Design Automation Conf. (DAC)

2020
[2020]

InProceedings of the 37th International Conference on Machine Learning

Transformers are RNNs: fast autoregressive transformers with linear attention. InProceedings of the 37th International Conference on Machine Learning
[2024]

BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation
[2025]

InProceedings of 2025 IEEE/ACM Design Automation Conference (DAC)

NetTAG: A Multimodal RTL-and-Layout-Aligned Netlist Foundation Model via Text-Attributed Graph. InProceedings of 2025 IEEE/ACM Design Automation Conference (DAC)

2025

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

https://www.tsmc.com/english/dedic atedFoundry/technology/logic/l_40nm

2008.TSMC 40nm LP process technology. https://www.tsmc.com/english/dedic atedFoundry/technology/logic/l_40nm

2008

[2] [2]

Design Compiler®RTL Synthesis

2021. Design Compiler®RTL Synthesis. https://www.synopsys.com/ implementation-and-signoff/rtl-synthesis-test/design-compiler-nxt.html

2021

[3] [3]

VCS®functional verification solution

2021. VCS®functional verification solution. https://www.synopsys.com/ verification/simulation/vcs.html

2021

[4] [4]

Xavier Bresson and Thomas Laurent. 2018. Residual Gated Graph ConvNets

2018

[5] [5]

Cadence. 2021. Innovus Implementation System

2021

[6] [6]

Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu

[7] [7]

Djork-Arné Clevert, Thomas Unterthiner, and Sepp Hochreiter. 2015. Fast and ac- curate deep network learning by exponential linear units (elus).arXiv:1511.07289 (2015)

arXiv 2015

[8] [8]

Jesse Davis and Mark Goadrich. 2006. The relationship between Precision-Recall and ROC curves. InProceedings of the 23rd International Conference on Machine Learning

2006