# Ams Sim

> Plan and run analog / mixed-signal simulations — DC op, AC, transient, Monte Carlo, corner sweeps, PSRR/CMRR, noise — using ngspice, Xyce, Spectre, or HSPICE, and triage the results. Use when the user says "SPICE", "ngspice", "Monte Carlo", "corner simulation", "analog sim", "AC analysis", "transient", "PSRR", "CMRR", "noise analysis".

- Skill: `vibeic/ams-sim` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx get-skill add vibeic/ams-sim`
- Raw SKILL.md: https://api.skillmd.com/api/skills/vibeic/ams-sim/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: vibeic (https://skillmd.com/u/vibeic)
- Updated: 2026-11-15
- Page: https://skillmd.com/skills/vibeic/ams-sim

---


# AMS Sim

Analog blocks must be verified across process corners, temperature, voltage, and mismatch. This skill plans the testbench matrix, generates the simulator decks, and triages results.

## When to use

- After `/analog-sizing` produces a schematic
- After `/analog-layout` + extraction (post-layout)
- When mismatching is the suspected failure mode (Monte Carlo)
- Mixed-signal co-simulation with a digital controller (ngspice + Verilog-A, or Spectre APS)

## Inputs

1. Schematic / netlist (`.sp`, `.cir`, `.scs`)
2. Device models from PDK (`.lib` / `.scs`)
3. Specs to verify (gain, BW, PSRR, CMRR, noise, slew, settling, offset)
4. Corner definition (FF/SS/TT/FS/SF × temp × supply)
5. Monte Carlo N (typical 500–5000 runs)

## Analysis matrix

| Analysis | Purpose |
|---|---|
| `.op` | DC operating point, bias validation |
| `.dc` | Sweeps (supply, input, parameter) |
| `.ac` | Gain, bandwidth, phase margin |
| `.tran` | Settling, slew, large-signal; sampled converter ENOB/SNR/SNDR via spectrum |
| `.noise` | Input-referred noise and small-signal analog SNR estimates |
| `.pss` / `.pac` / `.pnoise` | Periodic steady-state (for RF / SC circuits) |
| `.mc` | Monte Carlo mismatch + process |

The spec → required-analysis-subset mapping in this table is a frozen
lookup, **enforced by `programs/ams_analysis_select.py`** — do NOT pick
the analysis set by hand (two agents would otherwise pick different /
incomplete subsets for the same spec list). Run it on the block's
`spec.json` (or the project dir to scan `analog/*/spec.json`) and use
the emitted `analyses` set:

```bash
python3 programs/ams_analysis_select.py <block>/spec.json
# -> {"analyses": [".op", ".ac", ".tran", ".noise"], "mapping": [...]}
```

It always selects `.op` (bias validation) when ≥1 measurable spec is
present, adds `.mc` whenever any spec declares a mismatch / σ / yield
concern, and FAILs honestly (exit 1) when a spec carries no recognizable
measurable metric. Agent judgment is only needed to override an
ambiguous / custom spec key the table does not recognize.

ENOB and SINAD require a sampled dynamic measurement; declared converter
SNR/SNDR use transient capture and spectral processing as well. A small-signal
`.noise` result cannot establish these converter metrics. Separate physical
noise specifications still select `.noise`. See
[Analog Devices MT-003](https://www.analog.com/media/en/training-seminars/tutorials/MT-003.pdf),
pages 1 and 5–6. Stimulus, sampling and capture settings must be recorded with
the result; choosing an engineering measurement setting does not add a new
electrical acceptance requirement to the source specification.

## Workflow

1. Pick the analysis subset — **enforced by `programs/ams_analysis_select.py`** (see Analysis matrix)
2. Build deck with corner + MC sweeps
3. Run (ngspice / Xyce / Spectre / HSPICE)
4. Extract measurements via `.meas` statements — generated by `programs/analog_meas_from_spec_gen.py`
5. Read the native corner-sweep results for every required process, voltage and temperature point. Compare each measured specification with its original bound and report missing corners explicitly.
6. Review the native configuration and reports for every required corner and analysis role. Record missing libraries or unexecuted corners as NOT_MEASURED.
7. Hand off to re-sizing loop

## Output format

- `sim/<block>.sp` (or equivalent) — simulator deck
- `sim/<block>_results.md`:
  - Spec table with pass/fail per corner — populated from `corner_yield_vs_spec_check.py` (per-corner / per-spec PASS/FAIL re-derived from `spec.json` limits), not graded by eye
  - Read the native corner-sweep results for every required process, voltage and temperature point. Compare each measured specification with its original bound and report missing corners explicitly.
  - Sensitivity list (which device / parameter dominates) — agent judgment
  - Suggested fixes — agent judgment

## Tool prerequisites

Open source: ngspice (https://ngspice.sourceforge.io/), Xyce. Commercial: Cadence Spectre, Synopsys HSPICE, Silvaco SmartSpice.

## Technical basis

Pelgrom mismatch model, corner methodology from PDK, Monte Carlo with correlated process parameters. Mixed-signal co-sim per Verilog-AMS / Verilog-A standards.

## Handoff

- Failing corners → `/analog-sizing` (re-size)
- Layout-driven failures → `/analog-layout`
- Model issues → escalate to PDK team

## Captured rule

When model qualification has no declared deck, the producer writes one per
PDK device the delivered netlist instantiates, per process corner and
temperature of the analog corner grid: the device, terminal count and geometry
of its first instance, biased at the block's delivered rail value. The IC
expert reviews those decks instead of writing them, and declares only what does
not follow mechanically: a model library whose corner sections carry no
process-corner name (for example worst/best-case passives) gets its sections
named in `model_corner_sections`, otherwise that device stays at its nominal
section. A device with no declared role or defining library leaves the row
NOT_MEASURED with that reason. Read every simulator measurement; do not fill in
a result or add a performance limit to obtain a model-qualification PASS.

Keep a requested DC measurement point inside the simulator's actual sweep.
A descending decimal sweep can stop just above a nominal zero endpoint;
starting at the requested zero preserves the bias and measurement target.
Repair the project deck and rerun the native measurement, rather than
substituting the nearby trace value. A successful model/device check does not
measure converter accuracy, supply rejection or electrical noise budgets.

