acceptodds
Under review as a conference paper at ICLR 2027

HierRTL: Hierarchical Graph Representation Learning for Multi-Granularity RTL Performance Modeling

Abstract

Accurate power, performance, and area (PPA) assessment requires time-consuming synthesis and physical-design flows, limiting design-space exploration at the register-transfer level (RTL). Fast RTL prediction must capture both fine-grained hardware structure and design-wide context while preserving the association between logic and the state updates it supports. We propose HierRTL, a hierarchical Graph U-Net for joint design-level PPA and register-level timing prediction. HierRTL organizes RTL graphs into overlapping register cones that capture the data and control dependencies of individual next-state updates, with a directed cone graph modeling dependencies between registers. A U-shaped encoder–decoder aggregates information from nodes to cones and the design, then routes context back to nodes through the same cone memberships. This bidirectional hierarchy preserves shared-logic associations and produces contextualized node representations for predicting design-level area, power, worst negative slack, and total negative slack, together with register-level setup slack. To support local RTL optimization, we construct bounded local rewrites, verify them against the original RTL, and evaluate accepted implementations with a common flow. These equivalence-verified results define a soft opportunity target for each editable region. Starting from the jointly supervised predictor, HierRTL encodes the unmodified design once, summarizes each region only at readout, and learns four listwise rankings over the regions of the same design. Candidate implementations define supervision but are not required as model inputs. This connects multi-granularity RTL performance modeling with the practical task of prioritizing regions that are more likely to yield high optimization value before expensive search.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
RTL representation learning, hierarchical graph learning, graph neural networks, electronic design automation, PPA prediction, timing prediction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_yCMkOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

HierRTL builds a node–register-cone–design hierarchy from elaborated RTL and runs a Graph U-Net over it. The model is jointly trained on design-level PPA (area/power/WNS/TNS) and per-register setup slack (Stage I), then fine-tuned with a listwise loss to rank editable regions within a design using labels from LLM-generated, formally verified rewrites evaluated with a commercial flow (Stage II). Reported results are strong for Stage I: mean PPA NMAE 5.26 vs. 7.14 (RTL-Timer), and slack NMAE 3.59 vs. 4.75. For Stage II, within-design Spearman is 29.53 vs. 26.03 (GAT).

Critical strengths
  • Stage-I accuracy. Consistent gains on all four PPA targets and register slack over a broad baseline set on 22.8k designs with a design-level split.
  • Label pipeline. Only formally proven equivalent rewrites are kept and evaluated with a common commercial flow. The dataset may be reusable.
Critical weaknesses

My two main concerns are that 1) the technical novelty is limited relative to closely related work and 2) the Stage-II evidence does not yet support a strong region-ranking claim.

  • The “first” region-prioritization claim needs better positioning. Dr. RTL [1] already localizes optimization regions from register-level critical paths and uses a rewrite–synthesis–verification loop similar in spirit to the pipeline here. RTL-Timer also uses predicted slack to guide optimization. HierRTL may differ in learning an explicit ranking from rewrite outcomes, but this distinction needs to be made directly.

  • Register-cone hierarchy is not new. Register-centered subgraphs and register dependency graphs already appear in SNS v2, and register cones are used in CircuitFusion and TopoRTL. The main difference here seems to be the downward path over a fixed semantic hierarchy, which is close to a standard Graph U-Net decoder. The paper needs to isolate what is new beyond this combination.

  • Joint design- and register-level prediction is also not new. RTL-Timer predicts register slack and derives design-level timing quantities; MasterRTL aggregates register-endpoint paths into WNS/TNS; CircuitEncoder supports slack-, cone-, and PPA-level tasks with one encoder. This weakens the novelty of the multi-granularity supervision claim.

  • Overlapping cones need more analysis for area/power. HierRTL does not literally sum cone predictions, but shared logic enters multiple cone states before being routed back to nodes. The paper should explain why this does not bias area/power modeling. METIS is not a very relevant comparator for this issue.

  • Stage-II gains are modest relative to the reported variability. Spearman is 29.53 ± 1.00 vs. 26.03 ± 5.71 for GAT. Top-1 is 46.88 ± 1.11 vs. 45.61 ± 2.42 for CircuitFusion. With only 3 seeds, the evidence for a robust ranking gain is limited.

  • Key baselines are missing. Stage II has only about 3.3 regions per design on average, so a random Top-1 baseline is nontrivial. More importantly, there is no random, region-size, register-count, or worst-slack baseline. The last one is especially important because related work [1] already relies on timing criticality, and HierRTL itself predicts register slack.

  • Stage-II data may not reflect real optimization well. The paper does not clearly report the size/source distribution of the Stage-II subset. Also, labels are generated from rewrites produced by a single rewrite generator, so the target may partly measure where that generator can find improvements rather than the intrinsic optimization potential of a region.


[1] W. Fang, Y. Lu, S. Liu, J. Wang, Z. Guo, J. He, F. Tu, and Z. Xie. Dr. RTL: Autonomous Agentic RTL Optimization through Tool-Grounded Self-Improvement. arXiv:2604.14989, 2026.

Questions for authors

Q1. How is your region prioritization different from the critical-path localization in [1], which also maps register-level timing information back to RTL regions? How is it different from RTL-Timer-style guidance based directly on predicted slack? What does the learned ranking capture beyond a worst-slack heuristic?

Q2. Apart from using register cones as a fixed semantic hierarchy, what is new in the downward path compared with a standard Graph U-Net decoder? What can it capture that the bottom-up hierarchies in SNS v2, CircuitFusion, and TopoRTL cannot?

Q3. Shared logic can belong to multiple register cones. How does the model avoid bias from repeated cone membership for additive targets such as area and power? Also, the final design readout uses mean/max pooling. Where is design size encoded for size-dependent targets such as area?

Q4. The N column in Table 1 differs across baselines. Are all methods evaluated on the same test designs and trained with comparable budgets? Please also report standard deviations over seeds for Stage I.

Q5. How many editable regions does a typical Stage-II design have, and what fraction of region targets are zero or tied? How are tied targets handled in Spearman/Kendall and Top-1?

Q6. Fig. 2 mentions CEC/DSEC. Are sequential rewrites allowed? If so, when is sequential equivalence used, and what are the equivalence-check pass rates and retained candidates per region?

4 comments
AuthorsOpenReview ↗

Response for Reviewer yCMk(1/4)

Thank you for the comments on positioning and Stage-II evidence. We group related weaknesses and questions below. W and Q refer to the reviewer’s Weaknesses and Questions, respectively.


W1 & Q1. Positioning and timing criticality

The methods supply different decisions: Dr. RTL executes a rewrite loop, RTL-Timer supplies timing guidance, and HierRTL allocates regional exploration effort before candidate generation.

  • Dr. RTL[1]: Reads post-synthesis timing reports, maps critical paths to RTL, and uses its skill library to generate candidates. Synthesis and equivalence checking drive the next iteration. Region localization supports analysis and rewriting inside the optimizer.
  • RTL-Timer[2]: Predicts register-endpoint timing at the RTL stage, annotates HDL, and guides design and synthesis decisions such as path grouping and retiming.
  • HierRTL: Predicts regional opportunity rankings for Area, Power, WNS, and TNS from the original RTL and region definitions. Verified rewrite gains supervise where to spend a finite rewrite budget. Prediction requires graph processing and a model forward pass.

Our contribution is a regional prioritization task learned from verified rewrite gains. The model supplies exploration priorities to a downstream rewrite optimizer.

Worst slack measures current timing criticality, whereas learned ranking predicts gains discoverable by the specified rewrite flow. A critical region may have no feasible improving rewrite, while other regions may offer area or power gains. W6 tests this distinction directly.


W2 & Q2. Circuit-specific hierarchy and downward propagation

SNS v2[3], CircuitFusion[4], and TopoRTL[5] use register-centered representations. Following Graph U-Nets[6], HierRTL adds downward propagation along RTL dependencies and readout from decoded original nodes.

(1) Semantic hierarchy

  • Each cone contains the data/control dependencies of one register update, bounded by upstream registers.
  • Overlapping memberships preserve shared logic. The same memberships route information back to the original nodes.

(2) Downward propagation

  • Design context enters cone states and propagates over the register-dependency graph.
  • Each node combines its cones through per-node normalized learned weights, then gates this context with its local skip state.

(3) Shared node representation

  • Stage-I uses decoded nodes for design PPA and register slack.
  • Stage-II fine-tunes these states for regional ranking. Region memberships enter only at readout.

(4) Evidence

The matched-capacity control changes only the PPA readout:

PPA readout (NMAE×100↓)AreaPowerWNSTNSMean PPA
Design bottleneck5.6307.4896.1737.3976.672
Decoded cones5.6917.4355.9477.5226.649
HierRTL: decoded nodes4.3226.2135.0735.4335.260
  • Readout: Decoded nodes reduce mean PPA error by 21.17% relative to the design bottleneck.
  • Difficult structures: Against w/o cone & design, the full hierarchy reduces error by 19.2% on depth>11 designs and 27.3% on multi-register-feedback designs.

W3. Role of multi-granularity supervision

RTL-Timer[2], MasterRTL[7], and CircuitEncoder[8] already support related multi-granularity tasks. Our contribution is the jointly supervised node representation and its transfer to regional opportunity ranking.

(1) Complementary supervision on shared states

  • Four PPA heads read pooled decoded nodes to predict Area, Power, WNS, and TNS.
  • A shared register head predicts slack at nodes with valid endpoint labels.
  • Both losses update the same encoder and decoder, combining whole-design cost with local timing constraints.

(2) Transfer to Stage-II

  • Initialize the regional ranker with this circuit encoder, then fine-tune it using verified rewrite gains.
  • Regions query the decoded nodes only at readout, reusing the representation learned from both supervision levels.
AuthorsOpenReview ↗

Response for Reviewer yCMk(2/4)


W4 & Q3. Overlapping cones, PPA readout, and size information

(1) What cone overlap represents

A node may belong to several register cones. These cones provide contextual features, not separate area/power estimates to be added.

(2) Normalize incoming context

For a node vv with a nonempty cone set C(v)\mathcal C(v):

a_{vc} &= \frac{\langle Qh_v^{\mathrm{skip}},Kh_c^{\mathrm{dec}}\rangle}{\sqrt{d}}, \\ w_{vc} &= \frac{\sigma(a_{vc})}{\sum_{c'\in\mathcal C(v)}\sigma(a_{vc'})}, \\ q_v &= \sum_{c\in\mathcal C(v)}w_{vc}h_c^{\mathrm{dec}}. \end{aligned}$$ - $h_v^{\mathrm{skip}}$ is the local node state, and $h_c^{\mathrm{dec}}$ is a decoded cone state. - $Q,K$ are learned projections, $d$ is the hidden dimension, and $\sigma$ is sigmoid. - Nonnegative weights sum to one, making $q_v$ a weighted average rather than a sum that grows with cone memberships. **(3) Read each original node once** - Gate $q_v$ with the local skip state to produce one decoded feature per original node. - Area/Power heads read mean/max pooling over these unique nodes. Shared logic therefore occupies one readout position. This avoids explicit repeated cone contributions. Area/power remain learned outputs. METIS compares grouping semantics. **(4) Encode size before pooling** - **log(1+cone size) projection and count-scaled PNA aggregation** carry size information into the node states. - HierRTL has the lowest mean PPA error in all four timing-report register-name groups. - For the 19 Test designs with ≥1,000 register names, across-design mean PPA Spearman×100 is **88.34 for HierRTL**, versus 47.61 for RTL-Timer and 73.47 for TopoRTL. --- ### W5. Stage-II margins and variability We add complete graph and LLM-based comparisons, reporting means±sample SD. The added text encoders are Qwen3-Embedding[9] and CodeV[10]: | Method | Spearman×100↑ | Kendall×100↑ | | --- | ---: | ---: | | GCN | 22.97 ± 2.32 | 20.76 ± 1.87 | | GAT | 26.03 ± 5.71 | 23.55 ± 5.12 | | GIN | 24.15 ± 2.16 | 21.84 ± 1.78 | | RTL-Timer | 10.65 ± 3.87 | 10.00 ± 3.42 | | CircuitEncoder | 22.99 ± 2.55 | 21.06 ± 2.32 | | StructRTL | 6.49 ± 8.29 | 6.27 ± 7.76 | | TopoRTL | 13.91 ± 2.94 | 13.24 ± 2.59 | | CircuitFusion | 12.92 ± 2.31 | 12.25 ± 1.91 | > **Additional LLM-based methods and HierRTL** > > | Method | Spearman×100↑ | Kendall×100↑ | Parameters (M)↓ | Matmul GFLOPs/design↓ | Inference ms/design↓ | > |---|---:|---:|---:|---:|---:| > | Qwen3-Embedding 0.6B | 27.85 ± 0.43 | 25.85 ± 0.75 | 596.11 | 1,858.877 | 58.51 | > | Qwen3-Embedding 4B | 24.18 ± 1.77 | 22.37 ± 1.50 | 4,022.51 | 12,224.960 | 173.36 | > | Qwen3-Embedding 8B | 26.43 ± 1.18 | 24.63 ± 1.46 | 7,568.43 | 22,086.065 | 311.69 | > | CodeV-CL 7B | 19.38 ± 2.57 | 17.23 ± 2.30 | 6,608.54 | 27,812.254 | 355.79 | > | CodeV-DS 6.7B | 26.74 ± 1.85 | 24.45 ± 1.76 | 6,609.52 | 21,845.916 | 288.15 | > | CodeV-QC 7B | 20.36 ± 1.82 | 17.97 ± 2.14 | 7,071.62 | 20,326.649 | 285.39 | > | HierRTL | **29.53 ± 1.00** | **26.93 ± 0.91** | **1.47** | **0.187** | **9.13** | - **Task:** Learn regional priorities from verified rewrite gains, using only the original RTL at inference. - **Results:** HierRTL has the highest mean Spearman and Kendall, with fewer parameters and less computation than the text baselines. --- ### W6. Additional heuristics and classical EDA comparisons We first compare simple heuristics on the original Stage-II Test ranking set. HierRTL uses the Table 2 mean±sample SD: | Original Stage-II Test comparison | Spearman ×100 ↑ | Top-1 (%) ↑ | Normalized regret ×100 ↓ | |---|---:|---:|---:| | Uniform random, exact expectation | 0.00 | 29.82 | 63.45 | | Region register-group count | −3.05 | 29.88 | 62.93 | | Predicted worst endpoint slack | 6.79 | 35.71 | 58.02 | | HierRTL | **29.53 ± 1.00** | **46.88 ± 1.11** | **46.37 ± 1.51** | **Measured-report comparison** **(1) Designs** From the original Stage-II Test set (**202 designs, 676 regions**), include every design meeting both conditions for each target: - **Complete reports:** Every original region has a valid score, so both methods rank the same candidates. - **Defined ranking:** At least two regions and nonconstant opportunity labels. **(2) Ranking methods** - **Report rule, Area/Power:** Descending regional area/power share. Shared cells are split equally among their regions. - **Report rule, WNS/TNS:** Ascending worst setup slack among the region's register endpoints. - **Report sources:** Area/power use source-hierarchy-preserving synthesis with ≥95% traceable attribution. Slack uses the original implementation reports. - **HierRTL:** The original RTL graph and region memberships produce four opportunity scores, ranked in descending order. Training uses gains from equivalence-verified rewrites.
AuthorsOpenReview ↗

Response for Reviewer yCMk(3/4)

(3) Evaluation

Compare both methods against the same opportunity labels and average over designs for each target:

  • ρ: Within-design region-order agreement.
  • Top-1: Whether the selected region has the highest label.
ComparisonAREAPOWERWNSTNS
Report ruleArea sharePower shareWorst setup slackWorst setup slack
Report rule: ρ×100↑27.85−5.3914.5322.30
HierRTL: ρ×100↑31.7042.6522.5929.27
Report rule: Top-1 (%)↑44.4429.0353.6641.30
HierRTL: Top-1 (%)↑61.1161.2956.1050.00

HierRTL has higher mean ranking metrics for all four targets. Resource share and timing criticality do not fully capture rewrite opportunities.


W7 & Q5. Stage-II distribution, zero labels, and ties

(1) Region distribution

  • Full dataset: 1,919 designs and 6,093 regions, with 1–144 regions per design.
  • Test: 202 designs and 676 regions, with 1–41 regions per design.
  • Sources: DynamicRTL 758, RTLCoder 409, DeepRTL-v2 299, DeepCircuitX 254, and MG-Verilog 199 designs.
Editable regions per designFull dataset: designsTest: designs
186283
2–473478
5–923631
10–19646
≥20234

(2) Zero gains and ties

  • Zero label: No positive gain found for that target within the specified rewrite flow and budget.
  • Label scope: Opportunities depend on the generator. Beneficial regions can change across candidate pools.
  • Tied region: Another region in the same design has the same label, including zero-label ties. The table's columns can overlap.
TargetZero-gain regionsRegions with tied labels
Area397/676 (58.73%)319/676 (47.19%)
Power415/676 (61.39%)357/676 (52.81%)
WNS457/676 (67.60%)386/676 (57.10%)
TNS453/676 (67.01%)394/676 (58.28%)

Both columns cover all 676 Test regions.

(3) Evaluation

  • Evaluate designs with at least two regions and nonconstant labels.
  • Spearman: Average ranks for ties. Kendall: Tau-b.
  • Top-1: Any best-labeled region is correct, with tolerance 10⁻¹². Tied predictions use a fixed order. Constant predictions receive zero correlation.
  • Average designs equally within each target, then average the four targets. Listwise training skips single-region/constant-label groups and gives tied labels equal relevance.

Q4. Different N values and Stage-I variability

(1) Why N differs

Hardware and compute limits, method costs, and compatibility issues during adaptation and data processing leave some methods evaluated on smaller usable subsets. PPA label definitions are unchanged.

(2) Identical Test comparison

  • Take the prediction-coverage intersection of 14 methods: 2,018 common Test designs.
  • Reuse the original training-set normalization.
  • Mean PPA NMAE×100: RTL-Timer 6.311, TopoRTL 7.091, HierRTL 4.701.

The relative prediction advantage remains on identical Test inputs.

(3) Training budgets

  • Use the original design splits and Validation-based checkpoint selection, with budgets reflecting method costs.
  • External pretrained methods retain their initialization. Text encoders remain frozen while their heads are trained.

(4) Stage-I variability

Stage-I error, NMAE ×100 ↓Mean ± sample SD
Area4.322 ± 0.233
Power6.213 ± 0.372
WNS5.073 ± 0.079
TNS5.433 ± 0.105
Mean of four PPA targets5.260 ± 0.041
Register slack, macro3.586 ± 0.031

PPA covers 2,306 Test designs. Slack uses valid endpoint labels.


Q6. Sequential edits, equivalence checks, and candidate counts

(1) Allowed edits and equivalence

  • Edits may occur inside sequential process blocks under interface, clock/reset, and written-signal constraints.
  • Candidates must pass both CEC and DSEC. Unknown, timeout, and failed checks are rejected.

(2) Additional rewrite experiment

  • Formal checks: 1,087/1,199 candidates pass both proofs (90.66%), or 73.35% of all 1,482 planned attempts.
  • PPA: 1,085 completed measurement attempts, corresponding to 270 distinct implementations including the 17 originals.
  • Per region/pool: Two planned attempts, with 0–2 completed PPA measurements and mean 1.46.

(3) Original dataset

Historical exports retain 21,914 rewrites, of which 21,825 contribute to labels. Candidates per region range from 1 to 10, with mean 3.58.

AuthorsOpenReview ↗

Response for Reviewer yCMk(4/4)

References

[1] Wenji Fang, Yao Lu, Shang Liu, Jing Wang, Ziyan Guo, Junxian He, Fengbin Tu, and Zhiyao Xie. Dr. RTL: Autonomous agentic RTL optimization through tool-grounded self-improvement. arXiv preprint arXiv:2604.14989, 2026.

[2] Wenji Fang, Shang Liu, Hongce Zhang, and Zhiyao Xie. Annotating slack directly on your Verilog: Fine-grained RTL timing evaluation for early optimization. In Proceedings of the 61st ACM/IEEE Design Automation Conference, pp. 1–6, 2024.

[3] Ceyu Xu, Pragya Sharma, Tianshu Wang, and Lisa Wu Wills. Fast, robust and transferable prediction for hardware logic synthesis. In Proceedings of the 56th Annual IEEE/ACM International Symposium on Microarchitecture, pp. 167–179, 2023.

[4] Wenji Fang, Shang Liu, Jing Wang, and Zhiyao Xie. CircuitFusion: Multimodal circuit representation learning for agile chip design. In International Conference on Learning Representations, pp. 51523–51545, 2025.

[5] Mingyu Zhao, Xun He, Jiawei Liu, Jianwang Zhai, and Chuan Shi. Topology matters in RTL circuit representation learning. In The Fourteenth International Conference on Learning Representations (ICLR), 2026.

[6] Hongyang Gao and Shuiwang Ji. Graph U-Nets. In International Conference on Machine Learning, pp. 2083–2092. PMLR, 2019.

[7] Wenji Fang, Yao Lu, Shang Liu, Qijun Zhang, Ceyu Xu, Lisa Wu Wills, Hongce Zhang, and Zhiyao Xie. MasterRTL: A pre-synthesis PPA estimation framework for any RTL design. In IEEE/ACM International Conference on Computer Aided Design (ICCAD), pp. 1–9, 2023.

[8] Wenji Fang, Shang Liu, Hongce Zhang, and Zhiyao Xie. A self-supervised, pre-trained, and cross-stage-aligned circuit encoder provides a foundation for various design tasks. In Proceedings of the 30th Asia and South Pacific Design Automation Conference, pp. 505–512, 2025.

[9] Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. Qwen3 Embedding: Advancing text embedding and reranking through foundation models. arXiv preprint arXiv:2506.05176, 2025.

[10] Y. Zhao et al. CodeV: Empowering LLMs for Verilog generation through multi-level summarization. arXiv preprint arXiv:2407.10424, 2024.

Reviewer_8fvFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Accurate RTL prediction requires information from multiple granularities, while existing RTL prediction models usually lack the support of bidirectional information propagation across multiple granularities. To fill this gap, this paper proposes HierRTL, a hierarchical Graph U-Net model. It considers three RTL hierarchies: node, cone, and design. Here, we define a “cone” for each register, which traces data and control dependencies backward from its data input and stops at upstream registers. The U-Net encoder builds node embedding, cone embedding, and design embedding, while the decoder routes the context back from design to cone and to node. This bidirectional hierarchy can provide more context information and enable information flowing among long-range dependencies. HierRTL uses the final node embedding to predict register-level setup slack, and aggregates the final node embeddings to predict design-level prediction targets. Another contribution of this paper is that it can use the model to rank possible optimization regions in the RTL code, which could guide RTL code optimization. Experiment results demonstrate strong performance of the model.

Critical strengths

1, The idea is interesting. It is useful to consider the hierarchical structure of RTL code, and the bidirectional idea is inspiring. Besides, ranking possible optimization regions of RTL code might be a new task, and it can benefit practical applications.

2, In the Stage I experiment results, HierRTL clearly outperforms the baseline models. The Stage I experiment includes many baseline methods, including LLM-based methods, GNN-based methods, and EDA-based methods, and HierRTL outperforms them.

Critical weaknesses

1, The Stage II experiment results might be limited. (1) The best overall Spearman correlation is only 0.2953, which means that none of the methods perform very well. (2) It lacks LLM-based baseline methods, which exist in the Stage I experiments. (3) During training, the ranking of optimization regions depends on an LLM’s ability to generate possible rewrites. It would be better if the paper could provide some analysis of the LLM’s rewriting effectiveness and stability.

2, Some additional key experiments should be considered. (1) Currently, the design-level prediction comes from aggregation of node embeddings. Since the Graph U-Net can directly produce a design embedding at the design level, a natural idea is to do design-level prediction based on the design embeddings. This ablation can measure whether the downward path from design to cone to node is beneficial for design-level prediction. (2) It is meaningful to test the model performance on Stage I tasks after Stage II fine-tuning. This can measure whether Stage II fine-tuning degrades Stage I performance (please let me know if my understanding is wrong or if Stage I performance no longer matters after Stage II fine-tuning).

Questions for authors

Clarification requests:

1, Will you publicly share your code if accepted?

2, Why are some baseline methods evaluated on a different dataset size (different N in Table 1)?

3, How did you adapt the baseline methods to the Stage II experiments?

4 comments
AuthorsOpenReview ↗

Response for Reviewer 8fvF(1/2)

Thank you for the suggestions. We add LLM-based ranking comparisons, repeated rewriting, and readout controls. W and Q refer to the reviewer’s Weaknesses and Questions, respectively.


W1.1. Stage-II task and relative performance

(1) Predicting an unseen transformation

  • The input is the original RTL, but the target is improvement after an equivalent rewrite.
  • A local change may not improve the whole design. For example, shortening one critical path can expose another and limit WNS improvement.

(2) Learning from finite search

Labels are verified gains found within a finite rewrite budget. A zero label can mean that search did not find an improving candidate, making the ordering harder to infer.

(3) Relative performance

The absolute ranking performance leaves room for improvement. Under the same labels and design splits, HierRTL has the highest mean Spearman and Kendall among the graph and LLM methods in W1.2 & Q3.


W1.2 & Q3. Stage-II adaptation, ranking results, and efficiency

We adapt six Stage-I text encoders from Qwen3-Embedding[1] and CodeV[2].

(1) Text/LLM adaptation

  • Input: Encode the complete original RTL and region source separately with the same frozen encoder.
  • Features: Average and normalize chunk representations, then concatenate the design and region features.
  • Heads: Each model trains its own shared projection and four MLP heads for Area, Power, WNS, and TNS.

(2) Graph adaptation

  • Initialize from Stage-I. Encode the original graph, then combine attention-pooled regional nodes with global mean/max features for four ranking heads.
  • CircuitEncoder scatters cone context back to nodes. HierRTL fine-tunes its encoder and ranking heads jointly.

(3) Training and results

Use the same design splits, opportunity labels, and listwise loss. Select checkpoints by Validation ranking performance. Results are means±sample SD.

MethodSpearman×100↑Kendall×100↑
GCN22.97 ± 2.3220.76 ± 1.87
GAT26.03 ± 5.7123.55 ± 5.12
GIN24.15 ± 2.1621.84 ± 1.78
RTL-Timer10.65 ± 3.8710.00 ± 3.42
CircuitEncoder22.99 ± 2.5521.06 ± 2.32
StructRTL6.49 ± 8.296.27 ± 7.76
TopoRTL13.91 ± 2.9413.24 ± 2.59
CircuitFusion12.92 ± 2.3112.25 ± 1.91

Additional LLM-based methods and HierRTL

MethodSpearman×100↑Kendall×100↑Parameters (M)↓Matmul GFLOPs/design↓Inference ms/design↓
Qwen3-Embedding 0.6B27.85 ± 0.4325.85 ± 0.75596.111,858.87758.51
Qwen3-Embedding 4B24.18 ± 1.7722.37 ± 1.504,022.5112,224.960173.36
Qwen3-Embedding 8B26.43 ± 1.1824.63 ± 1.467,568.4322,086.065311.69
CodeV-CL 7B19.38 ± 2.5717.23 ± 2.306,608.5427,812.254355.79
CodeV-DS 6.7B26.74 ± 1.8524.45 ± 1.766,609.5221,845.916288.15
CodeV-QC 7B20.36 ± 1.8217.97 ± 2.147,071.6220,326.649285.39
HierRTL29.53 ± 1.0026.93 ± 0.911.470.1879.13

HierRTL retains the highest mean Spearman and Kendall, with about 1/405 the parameters and 1/6.4 the mean model-path inference time of the smallest text baseline.


W1.3. LLM rewriting effectiveness and stability

We generate three independent candidate pools for all 17 original Stage-II Test designs with at least eight editable regions: 247 regions, with two attempts per region per pool. Each attempt starts from the original RTL, and failures consume budget.

(1) Rewrite effectiveness.

StepAttempts
Planned attempts1,482
Candidate produced1,336
Edit contract passed1,199
Both CEC and DSEC proved1,087
PPA measurement completed1,085
Measured attempts changing RTL368
At least one PPA objective improved20

Improving candidates cover five designs and ten regions. Passing equivalence does not necessarily produce a PPA gain.

(2) Repeatability. All four PPA measurements agree at the saved precision for 96 repeatedly measured implementations. Beneficial regions vary across the three rewrite pools, showing that generation randomness also affects opportunity labels under a finite budget.

AuthorsOpenReview ↗

Response for Reviewer 8fvF(2/2)


W2.1. PPA readout at the design, cone, and node levels

We compare PPA readout from the upward design bottleneck, decoded cones, and decoded nodes, keeping the encoder, prediction heads, slack supervision, and parameter count unchanged. All variants retain the downward path for slack supervision.

PPA readout (NMAE×100↓)AreaPowerWNSTNSMean PPA
Design bottleneck5.6307.4896.1737.3976.672
Decoded cones5.6917.4355.9477.5226.649
HierRTL: decoded nodes4.3226.2135.0735.4335.260

Decoded-node readout reduces mean PPA error by 21.17% relative to the design bottleneck. This supports returning global context to local nodes before aggregation.


W2.2. Stage-I representations and Stage-II fine-tuning

  • Stage-I: Learn the initial circuit representation.
  • Stage-II: Fine-tune the encoder and ranking heads for regional opportunity ranking, selecting checkpoints by Validation ranking performance.
  • Deployment: Use the corresponding task-specific checkpoint for PPA prediction or region ranking.

Q1. Public code and release scope

  • Code: We plan to release model, baseline-adaptation, training, and evaluation code, with splits, configurations, and reproduction commands.
  • Restrictions: The complete training corpus and some commercial-EDA materials cannot be fully released under applicable license and redistribution terms.
  • Reproduction: Provide permitted processed data and checkpoints, data-source references, and preprocessing procedures for appropriately licensed environments.

Q2. Different N values in Table 1

(1) Why N differs

Hardware and compute limits, method costs, and compatibility issues during adaptation and data processing leave some methods evaluated on smaller usable subsets. PPA label definitions are unchanged.

(2) Identical Test comparison

  • Take the prediction-coverage intersection of 14 methods: 2,018 common Test designs.
  • Reuse the original training-set normalization.
  • Mean PPA NMAE×100: RTL-Timer 6.311, TopoRTL 7.091, HierRTL 4.701.

The relative prediction advantage remains on identical Test inputs.

References

[1] Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. Qwen3 Embedding: Advancing text embedding and reranking through foundation models. arXiv preprint arXiv:2506.05176, 2025.

[2] Y. Zhao et al. CodeV: Empowering LLMs for Verilog generation through multi-level summarization. arXiv preprint arXiv:2407.10424, 2024.

Thank you for your detailed reply and additional efforts! Your rebuttal deepened my understanding of the paper and addressed part of my concerns.

W1.1: It makes sense that "the absolute ranking performance leaves room for improvement." I think the stage-II task is interesting and might benefit future research.

W1.2: It is great to see HierRTL outperforms the LLM-based baseline models.

W1.3: Among 1085 rewrites that completed PPA measurement, only 20 improved at least one PPA objective. This might be a limitation of the Stage II contributions, as noted in my original review, because the ranking of optimization regions depends on an LLM’s ability to generate possible rewrites. Besides, the Stage II dataset is relatively small (only 202 for testing).

W2.1: The results clearly show that your design choice is correct.

W2.2: Thanks for the clarification that separate model checkpoints are used in two stages. Does it mean a single model cannot serve both stages well? It is better to explicitly introduce how models are used in two stages in the paper.

The rebuttal addressed part of my concerns, but as some concerns remain, I will keep the score at 2. I like the ideas proposed in this paper, and I believe it could become a good paper.

AuthorsOpenReview ↗

Follow-up response to Reviewer 8fvF

Thank you for your thoughtful follow-up and for recognizing the value of the task and the additional evidence. We would like to address your remaining concerns.

W1.3: Generator dependence and dataset scale

We agree that the limited rewrite success rate constrains the optimization benefit achievable with the current generator. The practical question addressed by Stage-II is where to allocate a limited rewrite budget given the available optimization tool. Generator dependence is therefore part of this decision problem. The observed success rate concerns candidate generation, while our ranking comparisons evaluate how well the model identifies promising regions under that same generation process. These are complementary aspects of an optimization system. As generators improve, their verified outcomes can provide updated supervision for this ranking task. Under the same labels and splits, HierRTL achieves the highest mean Spearman and Kendall among the evaluated graph and LLM baselines. This supports the value of the learned circuit representation for regional opportunity ranking.

Regarding scale, obtaining regional improvement labels requires candidate generation, formal equivalence checking, and PPA evaluation, making this dataset substantially more costly to construct than one with only design-level labels. As summarized in our response for reviewer 5kno(1/3), W1(1), in the dataset-size comparison table, our overall corpus contains 22,848 labeled RTL designs, the largest original-design count among the compared representation-learning datasets. The Stage-II subset requires this additional verification and measurement effort. We agree that expanding it would strengthen the evaluation, and view the present dataset and comparisons as a foundation for further research on this task.

W2.2: Practical use of the two checkpoints

Our intended applications do not require one checkpoint to retain both capabilities simultaneously. PPA prediction and regional opportunity ranking serve different purposes in the design workflow, so each can use its own task-specific checkpoint independently. Stage-I supplies both a predictor and the representation used to initialize Stage-II. After adaptation, the Stage-II checkpoint directly serves region ranking.

Maintaining and deploying these specialized checkpoints is practical given their low cost. As reported in W1.2 & Q3, the Stage-II model has 1.47M parameters and averages 9.13 ms per design for model-path inference, compared with 596.11M parameters and 58.51 ms for the smallest text baseline. HierRTL thus combines stronger mean ranking performance with about 405× fewer parameters and 6.4× faster model inference. In practice, we can independently deploy the appropriate checkpoint for each purpose without requiring the ranking checkpoint to retain the original regression capability. We will make this intended usage and the relationship between the two checkpoints explicit in the paper.

Thank you again for your careful assessment and constructive suggestions. Your feedback has helped us clarify both the practical role of Stage-II and how the models are intended to be used.

Reviewer_5knoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

HierRTL is a Graph U-Net for RTL. Yosys elaborates a design into a typed graph (operators, ports, constants, registers); for each register a backward traversal to upstream registers defines an overlapping "register cone", and register-to-register dependencies define a sparse cone graph. A PNA-based encoder aggregates node → cone → design; a decoder routes the design state back design → cone → node through attention-weighted cone memberships and gated skip fusion. Decoded nodes feed four design-level heads (area, power, WNS, TNS) and a per-register setup-slack head (Stage I). Stage II fine-tunes the encoder with a listwise objective to rank the editable regions of a design by "optimization opportunity", where the supervision is a softmax-weighted expectation of measured PPA gains from equivalence-verified local rewrites produced offline by an LLM agent (GPT-5.6 Luna) with synthesis feedback. At inference only the original graph and the region's node set are needed.

This is an interesting view on how RTL should be reasoned (with graph information embedded).

Critical strengths

Register cones with overlapping memberships and a deterministic node–cone–design structure are a natural inductive bias for RTL, and the U-shaped return path (routing design context back to nodes through the same memberships) is a clear.

Ranking editable regions by verified-rewrite opportunity, with candidates used only as offline supervision seems to be a reasonable practical way that mimics how RTL designers do the PPA optimization.

Critical weaknesses
  1. The designs are too small to motivate or demonstrate the work. Table 5: the median design has 54 RTL lines, 284 graph nodes and 4 registers (P95 = 21). Designs of this size go through Design Compiler synthesis-only, or even Fusion Compiler with physical synthesis, in minutes; the paper’s own labelling flow (DC + Innovus pre-route, App. A.1) is therefore not the “time-consuming” bottleneck the introduction invokes. At this scale a fast RTL predictor saves essentially nothing, and the results say little about the designs where prediction would matter. No results are reported as a function of design size and no design with hundreds of registers is included.

  2. The known hard cases for graph learning on circuits are untested. GNNs are known to struggle on circuit graphs with deep combinational logic (over-squashing across many hops) and with sequential feedback, where register-to-register dependencies form cycles and information must propagate around loops across time steps. These are exactly the cases the node → cone → design → node hierarchy is meant to address, yet with a median of 4 registers per design neither deep cones nor cyclic cone graphs occur often enough to be evaluated.

  3. The region-ranking task lacks a classical EDA baseline. All Stage-II comparisons are against learned models (Table 2). A designer prioritising regions today would use the synthesis reports they already have: rank regions by their area/power share, by the worst endpoint slack inside the region, or, since the designs are this small, partition the design by region and resynthesize each partition under different optimization settings to measure attainable gain directly.

  4. The hierarchy’s benefit for Stage-I regression is small and partly negative.

Questions for authors

Do not have questions to authors.

3 comments
AuthorsOpenReview ↗

Response for Reviewer 5kno(1/3)

Thank you for the comments. We address them with the additional experiments below.


W1. Design scale and practical motivation

(1) Coverage and size distribution.

Limited access to high-quality RTL training data with reliable implementation labels means that our dataset primarily comprises small modules. Its coverage of large industrial designs remains limited.

The graph's register nodes represent register cells or vector-register groups whose input dependencies define register cones, rather than individual synthesized FFs. To characterize implementation scale, we counted unique register names in the timing reports:

Timing-report register namesAll designsFraction of all designsTest designs
0–9918,96383.00%1,906
100–4993,35214.67%337
500–9992851.25%44
≥10002481.09%19

There are 3,885 designs with at least 100 reported register names. The corpus and Test maxima are 16,512 and 4,246, respectively.

Dataset sizes in related RTL representation-learning studies are:

Dataset / studyRTL designs
MasterRTL[1]90
RTL-Timer[2]21
SNS v2[3]89
CircuitEncoder[4]41
CircuitFusion[5]41
StructRTL[6]13,200
TopoRTL[7]115
HierRTL22,848

Our dataset contains the most labeled RTL designs among these datasets.

(2) Prediction by scale.

We group the same 2,306 Test designs by the register-name counts above and compare mean four-target PPA NMAE×100↓:

Method0–99 (N=1,906)100–499 (N=337)500–999 (N=44)≥1,000 (N=19)
Qwen3-E-0.6B5.81025.406130.456346.663
Qwen3-E-4B5.43624.341132.558284.263
Qwen3-E-8B5.83826.171126.953297.844
CodeV-CL-7B5.52523.808141.674272.705
CodeV-DS-6.7B5.80925.748138.845306.828
CodeV-QC-7B5.16121.084117.406325.648
GCN4.42521.649116.774376.263
GAT4.21420.933106.233346.867
PNA3.23811.89959.735152.682
RTL-Timer3.44612.72163.672148.082
TopoRTL3.25814.11470.148177.483
HierRTL2.4719.61740.736125.632

With limited training coverage of large designs, error increases with size for every method. HierRTL nevertheless retains the lowest error in all four groups.

We also measure PPA ordering across designs within each size group:

Mean PPA Spearman×100↑0–99100–499500–999≥1,000
RTL-Timer94.5291.4373.9147.61
TopoRTL94.5193.0181.1373.47
HierRTL96.6895.9491.6088.34

Although prediction errors grow on larger designs, HierRTL retains stronger across-design ordering than RTL-Timer and TopoRTL. The largest group contains 19 designs.

(3) Motivation and cost.

HierRTL predicts PPA through graph processing and model inference before DC synthesis and Innovus analysis, providing inexpensive feedback for candidate screening. The table includes graph processing and compares the same Test designs:

Flow (s)0–99 (N=1,906)100–499 (N=337)500–999 (N=44)≥1000 (N=19)
DC + Innovus pre-route evaluation26.03944.57768.071139.646
HierRTL: graph processing + model forward0.1210.2090.3891.728

Graph processing includes Yosys parsing and elaboration, hierarchy construction, data loading, and transfer. The model is already loaded.

The prediction flow offers a substantial time-cost advantage even for small and medium modules, supporting repeated candidate screening and indicating potential for use on larger circuits.

AuthorsOpenReview ↗

Response for Reviewer 5kno(2/3)


W2. Deep combinational logic and sequential feedback

(1) Structural groups

  • Combinational depth: Longest combinational path to a register data input, stopping at upstream registers. Thresholds 5 and 11 are the training median and 90th percentile.
  • Feedback: A closed register-dependency path involving multiple registers.

(2) Prediction results

Depth groups cover the full Test set. The feedback group can overlap them.

PPA NMAE×100↓Depth≤5 (N=1,421)Depth 6–11 (N=664)Depth>11 (N=221)Multi-register feedback (N=626)
GCN8.33715.65326.13811.576
GAT7.60214.56526.59311.875
GIN10.53615.43333.28914.775
w/o cone & design4.9946.94318.1196.961
HierRTL3.9075.03314.6425.061
  • GNN comparison: HierRTL has lower error than GCN, GAT, and GIN in both difficult groups.
  • Hierarchy ablation: Relative to w/o cone & design, error falls by 19.2% on depth>11 designs and 27.3% on multi-register-feedback designs.

W3. Report heuristics and regional opportunity ranking

We compare report heuristics with learned opportunity ranking.

(1) Designs

From the original Stage-II Test set (202 designs, 676 regions), include every design meeting both conditions for each target:

  • Complete reports: Every original region has a valid score, so both methods rank the same candidates.
  • Defined ranking: At least two regions and nonconstant opportunity labels.

(2) Ranking methods

  • Report rule, Area/Power: Descending regional area/power share. Shared cells are split equally among their regions.
  • Report rule, WNS/TNS: Ascending worst setup slack among the region's register endpoints.
  • Report sources: Area/power use source-hierarchy-preserving synthesis with ≥95% traceable attribution. Slack uses the original implementation reports.
  • HierRTL: The original RTL graph and region memberships produce four opportunity scores, ranked in descending order. Training uses gains from equivalence-verified rewrites.

(3) Evaluation

Compare both methods against the same opportunity labels and average over designs for each target:

  • ρ: Within-design region-order agreement.
  • Top-1: Whether the selected region has the highest label.
ComparisonAREAPOWERWNSTNS
Report ruleArea sharePower shareWorst setup slackWorst setup slack
Report rule: ρ×100↑27.85−5.3914.5322.30
HierRTL: ρ×100↑31.7042.6522.5929.27
Report rule: Top-1 (%)↑44.4429.0353.6641.30
HierRTL: Top-1 (%)↑61.1161.2956.1050.00

HierRTL has higher mean ranking metrics for all four targets. Resource share and timing criticality do not fully capture rewrite opportunities.

  • Regional resynthesis measures gains from changing EDA optimization settings.
  • Our task predicts opportunities in the specified LLM rewrite flow before candidate evaluation, allocating a finite rewrite budget.
AuthorsOpenReview ↗

Response for Reviewer 5kno(3/3)


W4. Overall hierarchy contribution and local-timing benefits of cones

We distinguish the full hierarchy from the incremental cone level.

(1) The full hierarchy reduces mean PPA error by 22.8%.

ModelPPA NMAE×100↓Register slack NMAE×100↓
w/o cone & design6.817.35
w/o cone5.673.87
HierRTL5.263.59

(2) Cone benefits concentrate on local timing, reducing error by approximately 30% for complex-dependency endpoints.

Cones yield a smaller increment for global PPA. Since cones organize logic around register-update dependencies, we also compare local slack:

Model (slack error↓)All Test: NMAE×100 (N=2,282)Cone nodes>40: NMAE×100 (N=238)Other upstream groups>5: NMAE×100 (N=206)Worst labeled endpoint: MAE/ns (N=2,282)
w/o cone3.8759.1145.5990.025268
HierRTL3.586 (−7.5%)6.200 (−32.0%)3.931 (−29.8%)0.022486 (−11.0%)

Both complex-dependency groups also improve on Validation, supporting cones for modeling local timing dependencies.

References

[1] Wenji Fang, Yao Lu, Shang Liu, Qijun Zhang, Ceyu Xu, Lisa Wu Wills, Hongce Zhang, and Zhiyao Xie. MasterRTL: A pre-synthesis PPA estimation framework for any RTL design. In IEEE/ACM International Conference on Computer Aided Design (ICCAD), pp. 1–9, 2023.

[2] Wenji Fang, Shang Liu, Hongce Zhang, and Zhiyao Xie. Annotating slack directly on your Verilog: Fine-grained RTL timing evaluation for early optimization. In Proceedings of the 61st ACM/IEEE Design Automation Conference, pp. 1–6, 2024.

[3] Ceyu Xu, Pragya Sharma, Tianshu Wang, and Lisa Wu Wills. Fast, robust and transferable prediction for hardware logic synthesis. In Proceedings of the 56th Annual IEEE/ACM International Symposium on Microarchitecture, pp. 167–179, 2023.

[4] Wenji Fang, Shang Liu, Hongce Zhang, and Zhiyao Xie. A self-supervised, pre-trained, and cross-stage-aligned circuit encoder provides a foundation for various design tasks. In Proceedings of the 30th Asia and South Pacific Design Automation Conference, pp. 505–512, 2025.

[5] Wenji Fang, Shang Liu, Jing Wang, and Zhiyao Xie. CircuitFusion: Multimodal circuit representation learning for agile chip design. In International Conference on Learning Representations, pp. 51523–51545, 2025.

[6] Yi Liu, Hongji Zhang, Yiwen Wang, Dimitris Tsaras, Lei Chen, Mingxuan Yuan, and Qiang Xu. Beyond tokens: Enhancing RTL quality estimation via structural graph learning. arXiv preprint arXiv:2508.18730, 2025.

[7] Mingyu Zhao, Xun He, Jiawei Liu, Jianwang Zhai, and Chuan Shi. Topology matters in RTL circuit representation learning. In The Fourteenth International Conference on Learning Representations (ICLR), 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.