- Rating
- 2
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper presents SMS-Sizing, a method that generates multiple feasible alternatives from a single design that has already been verified through simulation. SMS-Sizing explores alternatives by perturbing verified designs. It uses a pre-trained surrogate model to identify the next designs to simulate and to verify whether the proposed modifications to the search space are supported by available evidence. The experimental validation comprises 6 amplifier topologies assessing SMS-Sizing's ability to find solutions far from the initial design.
Critical strengths
S1. Feasibility is established by circuit simulation.
Surrogate predictions help identify candidate designs, but a design becomes a part of the set of feasible solutions only when circuit simulation confirms it satisfies all 8 criteria. Candidates predicted to be infeasible by surrogates remain eligible for simulation (Section 3.5). Empirical results back this decision. According to Table 24, screening of candidates based on average surrogate predictions would exclude ~25% (1289/5052) of simulator-proven feasible solutions.
S2. SMS-Sizing finds more widely different solutions.
AnchorOnly produces almost the same number of feasible designs as SMS-M (26.6 versus 26.2). Its solutions, however, remain closer to the initial design, and are less separated from one another, and have lower best worst-case constraint margins (Table 14). These results show that the complete method offers benefits beyond finding many nearby feasible solutions.
Critical weaknesses
W1. Contributions of individual elements are not well justified.
In the paper, improved search performance is attributed to surrogate scoring. However, the AnchorOnly model removes a few elements at once, including surrogate predictions, sensitivity scores, predicted feasibility masks, the served controller, gate choices, and box contractions (Appendix D.8). Moreover, selecting coordinates using sensitivity scores gives essentially the same feasible hit rate as uniform selection (19.63% versus 19.73%, Appendix D.10). Additionally, the separate NoContraction and NoGate controls retain similar distances from the initial design (Table 2). These results support the complete method, but leave the importance of its sensitivity analysis and evidence-gating mechanisms unclear. Appendix E also acknowledges that "the reported experiments do not compare against a quality-diversity baseline", leaving an important comparison missing for a contribution centered on producing multiple solutions.
W2. The necessity of multiple solutions is not sufficiently demonstrated.
The motivation for this work is that downstream layout-based decisions would benefit from having a collection of sizing vectors. However, the paper does not identify any specific downstream decisions or tasks that would benefit from an archive of solutions rather than selecting a single carefully optimized solution. The experiments fail to show any reductions in design effort or improvement in the results obtained from subsequent design phases. Furthermore, the computational cost is not clear. The surrogate model training uses 960000 records, and the 100 calls online budget fails to include pre-training, candidate pool inference, and control function calls (Section 4.1, Appendix C). Without a proper comparison of elapsed time and computational cost, it is not possible to prove whether the extra solutions justify the effort.
Questions for authors
Q1. Clarification. Contribution of the individual components.
Are there any experiments where only the surrogate score is varied? AnchorOnly achieves a similar feasible count, while NoContraction achieves a similar distance from the initial design. What is the evidence of further improvement due to the sensitivity analysis and evidence-gated updates?
Q2. Clarification. Comparison with other approaches aiming at finding multiple solutions.
What is the evidence of an advantage over other archive-based approaches when starting with the same initial feasible solution, simulator budget, and separation condition? As no such comparison is provided in Appendix E, how is the contribution evaluated independently of the drawbacks of those baselines tailored to other search purposes?
Q3. Clarification. Need for the diversity measures.
Which downstream analog-design decision benefits from the returned set of solutions rather than one carefully optimized design? On the simple tier, Figure 3 gives EEsizer a larger share of retained solutions when the minimum separation threshold is increased to 0.02 and 0.05. How should this result be weighed against SMS-M’s greater distance from the initial design ("Reach" in Table 1), and which measure better reflects practical usefulness?
Q4. Clarification. Computational effort.
What is the computational effort and time spent on SMS-Sizing (plus training of the offline surrogate model) relative to the baselines? Moreover, the up-front effort might be justified by the ability to reuse the same surrogate across several sizing sessions. At what number of such runs (and under which conditions) does this become worthwhile?