Declare geometry
Planar width, fin gated perimeter, and nanosheet gated perimeter are different conventions. Count and dimensions remain attached to the effective width.
DEVICE / COMPACT MODEL / NORMALIZATION
Make unlike numbers refuse comparison.
Device metrics become useful only after geometry, denominator, operating condition, extraction rule, model lineage, and evidence class are explicit.
ShapeGeometryBasisNormalizationStateBias + temperatureMethodDerivative ruleOriginModel lineageMeaningEvidence class
COMPARISON ANATOMY
Four declarations come before a ratio.
V137 uses original deterministic teaching fixtures. The numbers expose calculation and refusal behavior; they do not imitate a PDK, measured device population, or calibrated compact model.
Name the denominator
Total-device Ioff and Ioff per effective width are different metric contracts. A unit conversion cannot repair a missing geometry basis.
Lock the condition
Device, VGS, VDS, VBS, temperature, current sign, sweep direction, smoothing, and derivative rule define one comparable operating point.
Expose lineage
Model family, version, provenance, range, train residuals, holdout residuals, and evidence class travel with the result.
TWO USEFUL EQUATIONS
The equation never replaces its contract.
Equivalent dimensions do not guarantee equivalent meaning. Geometry, bias, sign, extraction, and evidence still have to match.
Ioff_norm = |Ioff_total| / Weffectivegm / Id = (dId / dVgs) / IdFor the same signed operating point, S/A and 1/V are dimensionally equivalent. Total-device A and A/um are not.
READING ORDER
Move from geometry to a visible stop.
- 01Geometry
Write the physical dimensions, body count, and effective-width or perimeter convention.
- 02Metric
Choose total-device or normalized quantity with its exact unit and denominator.
- 03Condition
Bind bias, temperature, sign, sweep, derivative, and model version.
- 04Evidence
Separate teaching, precomputed, measured, calibrated, and missing states.
- 05Stop
Refuse mixed contracts, unsupported corner probabilities, ranking, and signoff.
MODEL BOUNDARIES
Six distinctions prevent false confidence.
A synthetic line fit can teach slope, intercept, per-pad contact resistance, and sheet resistance. It is not a qualified process result.
DeltaV = I x Rseries is a one-pass illustration. A self-consistent compact-model correction requires a different solver and evidence contract.
Separating train and holdout residuals is necessary for model review, but a synthetic holdout does not establish physical correlation.
SS, TT, and FF name conditions. They are not probabilities, quantiles, or sigma points unless an independently validated mapping exists.
A finite sample has a seed, distribution meaning, sample count, mean, and sample deviation. It does not inherit a corner label.
Physical parameters, instance geometry, numerical controls, extracted metrics, and validation evidence answer different questions.
No PDK or proprietary model intake, arbitrary netlist execution, silicon calibration, predictive validation, device or foundry ranking, corner-to-probability conversion, qualification, disposition, tapeout, or signoff follows from this lesson or its deterministic lab.