ISSCC27: Comparator Atlas — schematic-to-layout decision characterization

[WLHsu0827]

Comparator Atlas — ISSCC 2027

Wei-Lun Hsu — National Tsing Hua University · Original code: MIT License

Notebook · Run in Colab · Reproduction instructions

Overview

A reproducible SKY130 comparator study connecting offset calibration, finite decision time and layout parasitics. It includes a netlist-checked MOS schematic, a nine-candidate comparison, a lower-energy control, actual layout/extraction, full nominal post-layout PVT results, vector figures and an interactive explanation of wrong versus unresolved decisions.

Results

Scope: RC timing and energy are results of the archived simulation decks; extracted-model physical fidelity remains unqualified. RC-versus-C differences are not isolated resistance effects or silicon predictions.

The original 49-condition, width-stressed schematic study uses local calibration: at 1 ns and sampled absolute input >=1 mV, original/selected correct counts are 310/392 and 372/392, with mean core energy 143.01 and 251.53 fJ. The post-selection lower-energy control gives 351/392 at 156.65 fJ.

The new nominal, code-zero study uses unchanged repaired 27-device geometry and Windows ngspice 47: five corners (TT/SS/FF/SF/FS), three supplies (1.62/1.8/1.95 V), three temperatures (-40/27/125 C), and four inputs (-10/-3/+3/+10 mV). Schematic and RC share 5 fF loads, common mode 0.5 VDD, a 10 ns clock, 50 ps edges and complementary 80%/20% decision rails.

Full 45-condition grid	Schematic	Extracted RC
Correct at 1 ns	180/180	156/180
Correct at prospectively declared 2 ns	180/180	180/180
Mean core energy, fJ/cycle	245.84	425.49

The other 24 RC points at 1 ns are unresolved, not wrong. Worst RC delay is 1.835 ns at FS / 1.62 V / -40 C / -3 mV. All 360 schematic/RC histories pass unchanged 10/5 ps numerical checks: 720 actual unique simulations, no refinement required. Mean per-point energy overhead is 73.45%; the ratio-of-population-means increase is separately 73.08%.

The earlier five-condition ngspice-42 pilot remains distinct: its original 1 ns RC gate failed (12/20); 20/20 at 2 ns was post-hoc. Historical DRC/LVS and negative-control evidence is retained; extraction was not repeated for the new grid. Five conditions were previously observed, so this is not a blinded external test. Recorded execution continuations and the original timing-conformance failure remain in the audit; completed data qualification is not an original-budget timing claim.

Reproduction and scope

Run all notebook cells to check tables, regenerate plots and remeasure ten representative full-PVT traces—not all 720 raw traces. Three vector PDF/SVG figure masters have 600 dpi PNG companions. Existing waveform teaching examples are illustrative, not extra coverage.

The isolated clean-start reproducer accepts existing local tool/model paths and fresh output. Its four-run schematic/RC smoke passed numerical and same-point comparisons; a fresh full 720-run replay through that entrypoint was not performed. REPRODUCIBILITY.md gives commands, pinned Windows runtime/model identities, evidence locations and history. No paid API, commercial EDA license or cloud runner is required; Colab is optional and free-tier limits apply.

Core energy excludes external drivers and calibration infrastructure. No silicon, yield, noise probability, foundry signoff, new topology or continuous-input guarantee is claimed. A 3.5 ns deadline was not evaluated in the full-grid RC study. Original code is MIT licensed; third-party notices and references are preserved. GitHub Copilot assistance is disclosed, and the author remains responsible for the work. Only ISSCC27/submitted_notebooks/comparator_atlas/ changes; no root workflow or other entry is modified.

Verification and repository checks

For revision 8011aba, local Windows/Python 3.10.18 review passed 127 Python tests, 13 JavaScript tests, official nbqa lint and the complete 26-cell notebook. Package hashes and local links were verified. New PVT and four-run clean-start Windows evidence are separately audited, not certified.

No independent Linux review or manual cloud CI was run for this head. The earlier Linux review applies only to 0cfc795, not this new dataset or revision. Inherited root workflows fail to locate the nested notebook with literal **/*.ipynb; they remain unchanged. Automatic root checks are separate from local entry review; maintainer handling is not a policy-bypass request.

[WLHsu0827]

Hello Code-a-Chip organizers,

Could you please confirm that PR #195 is in the ISSCC 2027 submission queue and advise whether any additional entry-specific action is required?

I found a deadline discrepancy in the current official announcements: the GitHub program page says October 9, 2026, 11:59 AM Pacific, while the SSCS award page says October 31, 2026, 11:59 AM Pacific. Which date applies? I am treating the earlier October 9 date as the working deadline until confirmed.

The notebook openly discloses GitHub Copilot assistance. Please also let me know whether this program has any additional AI-use or disclosure requirements beyond the published rules.

All submission files are under ISSCC27/submitted_notebooks/comparator_atlas/, with an open-source license, references and reproducibility instructions. The entry-specific Linux review passes. The inherited root checks fail on the literal **/*.ipynb path; I have not changed root workflows in this PR and am not requesting a policy bypass.

Thank you,
Wei-Lun Hsu
National Tsing Hua University