- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes GIF for IR-drop prediction, combining layout feature maps and gate-level netlist representations within a conditional diffusion model. Experiments on CircuitNet-N28 and N14 evaluate prediction accuracy, hotspot detection, and uncertainty estimated from multiple generated samples.
Critical strengths
- GIF adapts established multimodal conditional diffusion ideas to combine complementary layout and netlist information for IR-drop prediction. This offers a useful application framework that produces both spatial predictions and uncertainty estimates.
Critical weaknesses
- The paper does not sufficiently distinguish its contribution from related topology–geometry encoding and multimodal fusion methods, such as DeepLayout[1] and LMM-IR[2]. Its empirical competitiveness also remains unclear without comparisons against more recent IR-drop predictors.
[1] DeepLayout: Learning Neural Representations of Circuit Placement Layout. ICML 2025.
[2] LMM-IR: Large-Scale Netlist-Aware Multimodal Framework for Static IR-Drop Prediction. DAC 2025.
Questions for authors
- Appendix Tables 12–13 report GIF’s sampling times. How does its inference latency, including four-sample averaging, compare with the regression baselines under matched hardware and input resolution?
- Beyond the graph/no-graph comparisons in Table 4, can the authors isolate the contribution of logical connectivity by retaining node features and model capacity while removing or randomizing edges?
3 comments
Thank you for the detailed response, the clarification regarding DeepLayout and LMM-IR, and the additional experiments. Two concerns remain:
-
Recent baselines. My concern in W1 was the lack of direct comparisons with representative methods published in 2025–2026 at major EDA or AI venues that address the same task or can reasonably be adapted using comparable inputs. Another reviewer 7htC raised a related concern about the baseline coverage.
-
Contribution of logical connectivity. Rewiring the edges at inference leaves Pearson unchanged, and removing them improves SSIM with little change in Pearson. The training comparison with and without graph tokens measures the effect of the entire graph branch. A matched training comparison using real versus randomized edges, with node features and architecture held fixed, would better isolate the contribution of logical connectivity.
Given these remaining concerns, I maintain my rating of 2.
Thank you for for reading and the follow-up. We answer the concerns from the follow up below.
1. "Direct comparisons with representative methods published in 2025-2026 ... that address the same task or can reasonably be adapted using comparable inputs."
The recent methods solve a different problem or give a different output:
| Method | What it estimates | Power-grid input | Logical netlist | Output |
|---|---|---|---|---|
| LMM-IR (DAC 2025) | static IR drop | SPICE netlist of the power grid | no | one IR drop map |
| PDNNet (2024) | dynamic IR drop | graph of the power grid | no | one IR drop map |
| DuST (DATE 2025) | dynamic IR drop | none | no | one IR drop map |
| GIF | dynamic IR drop | none | yes | IR drop map and uncertainty map |
- LMM-IR solves static IR drop and needs the power-grid SPICE netlist. It is built for a different problem and a different input.
- DuST solves the same task and uses the same training and test layouts as Table 3. However, its own evaluation code gives MAVIREC an SSIM of 0.761, while Table 3 lists 0.68. Its numbers are therefore on a different scale (Section 4).
- GIF has the highest SSIM (0.79) and Pearson (0.95) among the methods in Table 3, with one sample. It is the only method that gives an uncertainty map; this uncertainty is 2.0x higher on missed hotspots.
2. "A matched training comparison using real versus randomized edges ... would better isolate the contribution of logical connectivity."
GIF's claim is about the netlist as a whole, not about single edges. By design, the graph tokens are one summary of the whole netlist (mean pooling, Section 3). The test that matches this claim is the matched training comparison we ran, with the same architecture and the same features, with and without graph tokens, five training runs:
- SSIM: 0.765 → 0.780 with graph tokens.
- SSIM is higher on 90% of the test layouts.
- Pearson varies 9x less across seeds.
We hope these clarifications address your concerns. We would be grateful if you would consider updating your score, and we are happy to answer any further questions during the discussion.
We thank the reviewer for the careful reading and the valuable comments. Below are our responses to each concern; we ran both suggested experiments. All changes in the revised PDF are marked in blue.
1. "The paper does not sufficiently distinguish its contribution from ... DeepLayout and LMM-IR." These two works solve different problems from GIF:
We added both papers to Table 1 and the related work.
2. "Its empirical competitiveness also remains unclear without comparisons against more recent IR-drop predictors." The paper already compares GIF with recent IR drop methods; in both comparisons, GIF has the highest Pearson and is the only method with an uncertainty map. Table 3 (the same test layouts as PDNNet; GIF uses one sample):
Table 5 (regressors trained on the same 34-channel input):
For the recent DuST (DATE 2025), its own evaluation code gives MAVIREC an SSIM of 0.761 instead of 0.68, so its numbers are on a different scale (Section 4), and it gives no uncertainty map.
3. Q1: "How does its inference latency, including four-sample averaging, compare with the regression baselines?" One RTX A6000, same input, one test layout:
4. Q2: "Isolate the contribution of logical connectivity by retaining node features and model capacity while removing or randomizing edges." GIF is robust to edge changes, by design. Test on the trained N28 model (96 test layouts, same noise in all rows except the last):
The graph tokens are one summary of the whole netlist (mean pooling), so single edge changes move the IR drop map 7x to 24x less than the sampling noise does. The contribution of the netlist is learned in training: models trained with graph tokens have higher SSIM (0.780 vs. 0.765) and 9x more stable Pearson across seeds (Appendix).