- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes CritiCell, a multiple instance learning framework that infers cell-level timing criticality from path-level labels and guides two placement optimization workflows.
Critical strengths
-
Addresses a practically relevant problem without requiring cell-level annotations.
-
Two placement applications, feature ablations, and heuristic comparisons support its effectiveness.
Critical weaknesses
-
The max-pooling assumption may not adequately capture the cumulative nature of path delays. Placement improvements alone do not establish accurate cell-level attribution.
-
Main results aggregate training, validation, and unseen circuits while using circuit-specific hyperparameters, limiting the evidence for generalization.
Questions for authors
-
Can controlled local-movement experiments with equal optimization budgets verify that highly scored cells offer greater timing improvement potential?
-
How does the method perform across multiple random seeds on unseen circuits with fixed hyperparameters?
5 comments
Response to Reviewer vNdj — Part 2
Weakness 1
Thank you for raising these points. STA accounts for cumulative delays along each path when computing slack. In CritiCell, max pooling aggregates learned cell scores to predict the binary path label (Eq. 4). This lets the model learn from path-level supervision under the assumption that a timing-critical path contains at least one bottleneck cell. Our training filter excludes long violating paths to reduce cases where delay accumulated across many stages dominates the path label (Section 3.3).
The controlled single-cell moves in our response to Question 1 provide direct cell-level evidence. Under the same local search budget, cells with high CritiCell scores yielded larger mean path-delay reductions than either baseline on all three unseen circuits.
Weakness 2
Thank you for raising this point. We report the mean per-circuit WNS / TNS gains over the timing-aware baselines in Tables 1–2, grouped by the model development split in Appendix B.3.
| Comparison | Training | Validation | Unseen test |
|---|---|---|---|
| Ours (Clust.) vs. TritonPart | +11.82% / +12.40% | −6.52% / +7.11% | +3.16% / +12.06% |
| Ours (TDP) vs. Efficient-TDP | +5.46% / +8.79% | −0.98% / +15.48% | +6.08% / +9.05% |
Both placement flows improve WNS and TNS on average on circuits held out from model development. To assess performance with a common parameter-selection procedure, our response to Question 2 reports a multi-seed evaluation in which a shared rule sets hyperparameters. In comparisons using the same placement seed for both methods, Enhanced Efficient-TDP improved WNS by 6.38% and TNS by 6.43% on average over the original Efficient-TDP.
We have added the results by data split to Appendix B.3 of the revised paper.
Request for Feedback on Our Responses
Dear Reviewer vNdj,
We would appreciate it if you could review our responses when you have a chance. We hope they address your concerns and clarify our work. If you have any further questions or concerns, please let us know. We would be happy to continue the discussion and provide additional clarification.
Thank you again for your time and thoughtful feedback.
Best regards,
The Authors
Response to Reviewer vNdj — Part 1
Dear reviewer vNdj,
Thank you for your insightful comments. We sincerely hope our rebuttal could adequately address your concerns. If so, we would deeply appreciate it if you could consider raising your score. If not, please let us know your further concerns, and we will continue actively responding to your comments.
Question 1
Thank you for suggesting this test. For each of the three unseen circuits, we sampled 16 timing-violating paths from a legalized DREAMPlace placement. On each path, we selected the cell with the highest CritiCell score, the cell with the highest local-delay score, and a uniformly random cell.
Each selected cell received the same search budget of eight predefined displacement candidates within ten row heights, plus a no-op. We applied each legal move independently from the same baseline placement, recomputed placement parasitics, and reran STA. For each cell, we recorded the largest delay reduction on the same fixed data path among the tested moves.
Mean best path-delay reduction across 16 paths per circuit (ps)
Under the same local search budget, cells selected by CritiCell yielded larger mean path-delay reductions than either baseline on all three unseen circuits, showing that cells with high CritiCell scores offer greater timing improvement potential.
We have added this equal-budget local-movement experiment to Section 4.3 of the revised paper.
Question 2
Thank you for asking about sensitivity to random seeds. We evaluated Enhanced Efficient-TDP with three placement seeds on each of the three circuits unseen by the trained model. Since timing conditions vary across circuits, we used a shared empirical rule to set the two attraction multipliers based on the initial STA feedback for each run, then kept those values fixed for the remainder of placement.
Let N and H denote the counts of repeated non-critical and critical pin-pair occurrences, respectively, divided by the number of movable cells.
αnon=clip(N−0.4,0,2),αcrit=clip(0.04/H,1,2),with αcrit=1 when H=0.
The table reports WNS / TNS improvements for Enhanced Efficient-TDP over the original Efficient-TDP on the same circuit and placement seed.
Averaged equally across circuits, WNS and TNS improved by 6.38% and 6.43%, respectively, with both improving in 7 of 9 paired comparisons. These results demonstrate the robustness of our method.
We have added the shared parameter rule and the multi-seed results to Appendix C.9 of the revised paper.