Comparator

StrongArm Comparator Skill

Dependencies: ngspice on PATH, Python 3 + numpy, matplotlib, scipy.

Layout

comparator/
├── SKILL.md
├── scripts/                        ← run from here (cd comparator/scripts)
│   ├── ngspice_common.py           — paths, ngspice runner, parser, template renderer
│   ├── comparator_common.py        — circuit params, DUT rendering, τ/Tcmp/noise helpers
│   │
│   ├── simulate_tran_strongarm_wave.py   — backend: 3-cycle waveform + τ extraction
│   ├── simulate_tran_strongarm_noise.py  — backend: noise (probit) + power + Tcmp
│   ├── simulate_tran_strongarm_ramp.py   — backend: ramp transfer curve
│   ├── plot_tran_strongarm_wave.py       — plot: all internal nodes
│   ├── plot_tran_strongarm_noise.py      — plot: CDF fit + statistics
│   ├── plot_tran_strongarm_ramp.py       — plot: transfer curve
│   │
│   ├── run_tran_strongarm_comp.py   ← MAIN entry point (wave + noise + ramp in parallel)
│   ├── run_tran_strongarm_wave.py   — standalone: waveform only
│   ├── run_tran_strongarm_noise.py  — standalone: noise + FOM only
│   ├── run_tran_strongarm_ramp.py   — standalone: ramp only
│   │
│   ├── run_tcmp_sweep_plot.py       — sweep W_tail/inp/lat_n/lat_p → Tcmp plots
│   ├── run_tau_sweep_plot.py        — sweep W_lat_n/lat_p/inp → τ plots
│   ├── run_power_sweep_plot.py      — sweep W_tail/inp/lat_n/lat_p → power plots
│   │
│   ├── run_sweep_input_amplitude.py — Vin_diff → Tcmp curve
│   ├── run_sweep_vcm.py             — VCM → σ_n
│   ├── run_sweep_noise_bw.py        — noise BW → σ_n
│   ├── run_sweep_k_valid.py         — σ_n vs input amplitude (validation)
│   ├── run_compare_wave.py          — StrongArm vs Miyahara waveform comparison
│   └── diag_high_vcm.py            — diagnostic: behaviour at high VCM
│
├── assets/
│   ├── netlist/                    ← SPICE templates (*.cir.tmpl)
│   └── models/ptm45hp.lib          ← PTM 45nm HP BSIM4 model
│
└── references/                     ← read when user asks about theory/noise/offset
    ├── 01_theory.md
    ├── 02_speed.md
    ├── 03_noise.md
    └── 04_offset.md

All generated files (logs, plots, netlists) go to WORK/ at the repo root — never inside the skill package. Override with env-var ANALOG_WORK_DIR.

Task → Script Decision Guide

Running

cd comparator/scripts
python run_tran_strongarm_comp.py

Runs 3 simulations in parallel (~10s on a modern PC):

1. Waveform — Vin=+1mV, 3 cycles, all internal nodes → τ extracted
2. Noise — 1000 cycles with trnoise → P(1) → σ_n via probit; also measures power and Tcmp
3. Ramp — Vin ramps −2mV→+2mV, 100 cycles → transfer curve

Key outputs in WORK/:

• plots/strongarm_waveform.png
• plots/strongarm_noise.png
• plots/strongarm_ramp.png
• logs/fom_report.txt — σ_n, P_avg, Tcmp, FOM1, FOM2

Circuit Topology

StrongArm comparator, 45nm PTM HP, VDD=1.0V, CLK=1GHz, VCM=0.5V.

Transistor  Type   Gate  Drain  Source  Role
──────────────────────────────────────────────────────────
M0          NMOS   CLK   VS     GND     Tail (evaluation switch)
M1          NMOS   INP   VXP    VS      Input pair (+)
M2          NMOS   INN   VXN    VS      Input pair (−)
M3          NMOS   VLN   VLP    VXP     NMOS latch (stacked on M1)
M4          NMOS   VLP   VLN    VXN     NMOS latch (stacked on M2)
M5          PMOS   VLN   VLP    VDD     PMOS latch (cross-coupled)
M6          PMOS   VLP   VLN    VDD     PMOS latch (cross-coupled)
M7–M10      PMOS   CLK   VX/VL  VDD     Reset (pull to VDD when CLK=0)
M11–M14     CMOS   VLP/VLN  OUTP/OUTN  Output inverters

Critical connection: M3 source = VXP (NOT GND). As VXP drops during integration, Vgs_M3 = VLN − VXP increases → M3 turns on when VXP < VDD − Vth. Current path: GND → M0 → M1 → VXP → M3 → VLP.

Output polarity: OUTP = NOT(VLP) = 1 when INP > INN.

Transistor Sizes (all L = 45 nm)

Defined in comparator_common.py → W = dict(...):

Noise Model

Two independent trnoise sources at INP/INN (band-limited white noise):

# comparator_common.py
NOISE_BW = 100e9        # Hz  (100 GHz — decorrelates between 1 GHz cycles)
NOISE_NT = 1/(2*NOISE_BW)  # 5 ps  update interval
NOISE_NA = 300e-6       # V   amplitude per side

Noise is a stimulus parameter only; true σ_n is extracted from P(1) statistics via the probit method: σ_n = Vin_fixed / norm.ppf(P1) with Vin_fixed = 0.35 mV.

Output & File Conventions

• Never open/pop-up figures — only save PNGs via fig.savefig() + plt.close().
• Netlists → WORK/netlists/dut/ (DUT) and WORK/netlists/testbench/ (testbenches).
• Comparison plots: max 3 vertically stacked subplots. Group related signals:
  1. CLK + INP/INN
  2. Latch nodes VLP/VLN
  3. Output OUTP/OUTN

References