gm/ID Characterization and Design Skill

Important — do not modify skill files during normal use. All code edits, new scripts, plots, and simulation outputs should go into the user's project working directory (outside .claude/), not into this skill folder. Only modify the skill assets (assets/, SKILL.md, references/) when the user explicitly asks to improve or extend the skill itself.

Dependency: ngspice skill (netlist execution, wrdata parsing). Model files are built in — the transistor-models skill is not required.

Asset Files

assets/
├── simulate_gmoverid.py   — ngspice simulation engine, data extraction, analytical caps, MODEL_INFO registry
├── plot_gmoverid.py       — all matplotlib plotting functions
├── run_gmoverid.py        — 180 nm single-node orchestration (NMOS/PMOS + channel-length sweep)
├── run_multinode.py       — multi-node orchestration (45/22 nm HP)
├── design_gmoverid.py     — GmIdTable lookup/sizing API + print_op()
├── models/                — built-in PTM model files (ready to use)
│   ├── nmos180.lib        — 180 nm BSIM3v3 (VDD = 1.8 V)
│   ├── pmos180.lib
│   ├── nmos45hp.lib       — 45 nm HP BSIM4 (VDD = 1.0 V)
│   ├── pmos45hp.lib
│   ├── nmos22hp.lib       — 22 nm HP BSIM4 (VDD = 0.8 V)
│   └── pmos22hp.lib
└── netlist/
    ├── gmoverid_vgs.cir.tmpl       — NMOS Vgs sweep (fixed Vds)
    ├── gmoverid_vds.cir.tmpl       — NMOS Vds sweep (fixed Vgs)
    ├── gmoverid_pmos_vsg.cir.tmpl  — PMOS |Vsg| sweep (fixed |Vsd|)
    └── gmoverid_pmos_vsd.cir.tmpl  — PMOS |Vsd| sweep (fixed |Vsg|)

Built-in models are from PTM — Arizona State University (ptm.asu.edu), free for academic research. When citing, use: W. Zhao and Y. Cao, "New Generation of Predictive Technology Model for Sub-45 nm Early Design Exploration," IEEE Trans. Electron Devices, vol. 53, no. 11, pp. 2816–2823, Nov. 2006. doi: 10.1109/TED.2006.884077. For nodes beyond the three built-in ones, install the transistor-models skill (full PTM library, requires manual MODEL_INFO configuration) or download from mec.umn.edu/ptm and copy the .lib file into the project's models/ directory.

Workflow 1: Characterization (Simulation + Plotting)

Deployment

1. Copy all assets (including the models/ subdirectory) to <project>/
2. Create empty directories plots/ and logs/
3. Run python run_gmoverid.py (180 nm) or python run_multinode.py (HP multi-node)

For additional nodes, simply copy the extra .lib file into <project>/models/.

All paths resolve automatically via Path(__file__).resolve().parent — no path configuration needed.

Three Standard Plot Sets (per model)

gm/ID four-quadrant layout (plot_main, 2×2):

(0,0) gm/Id [V⁻¹] vs Vov     | (0,1) Id/W [uA/um] vs gm/Id  (log Y)
(1,0) fT [GHz]    vs gm/Id   | (1,1) gm·ro        vs gm/Id

• All four panels use Vds as the curve parameter (VDS_LIST)
• gm/Id X-axis: xlim = [4, 24], one grid division per 2 V⁻¹
• Panels (0,1), (1,0), (1,1) apply _fall_mask(gmid) before plotting — the gm/ID array is double-valued (same gm/ID appears on the subthreshold rising side and the inversion falling side); only the falling (inversion) side is kept to prevent curve fold-back. Panel (0,0) is exempt because its X-axis is Vov, which is monotonic.

Adding a New Technology Node

1. Copy the model .lib into models/ (download from mec.umn.edu/ptm or install the transistor-models skill)
2. Add an entry to MODEL_INFO in simulate_gmoverid.py (see conventions.md §3)
3. Add an entry to NODE_CFG in run_multinode.py

Plotting Conventions (required when creating new figures)

When generating figures for a new project or plotting for a design report, follow the style of the existing figures produced by this skill.

Format requirements:

• Never call plt.show(); always use fig.savefig(path, dpi=150, bbox_inches='tight') and save to plots/
• Colors and line styles: cycle through COLORS/LSTYLE from plot_gmoverid.py in order (blue solid, orange dashed, green dash-dot, purple dotted)
• X-axis for gm/ID: always xlim = [4, 24] with one tick per 2 V⁻¹; Y-axis ylim(bottom=0)
• Fold-back prevention: whenever gm/ID is the X-axis, prepend _fall_mask(gmid) to the boolean mask. The gm/ID vs Vgs curve peaks and then descends — both sides can fall within [4, 24], causing matplotlib to draw a fold-back. _fall_mask discards the subthreshold rising side (before the peak); only the inversion falling side is plotted. This rule applies to plot_main panels (0,1)/(1,0)/(1,1) and to the equivalent panels in plot_comparison.
• Titles and axis labels: use ASCII + LaTeX (e.g. $g_m/I_D$); do not write Chinese characters directly in matplotlib labels
• µ (U+00B5) fails to render in some terminals and fonts — always use u in axis labels (e.g. uA/um, W=10um) or LaTeX $\\mu$; use Unicode µ only when the user explicitly requests it

Chinese font warning: matplotlib does not include Chinese fonts by default; Chinese characters render as boxes (□□). If Chinese is needed, add at the top of the script:

import matplotlib
matplotlib.rcParams['font.sans-serif'] = ['Microsoft YaHei', 'SimHei', 'DejaVu Sans']
matplotlib.rcParams['axes.unicode_minus'] = False

On Windows prefer Microsoft YaHei; on Linux/macOS verify the font is installed, otherwise boxes still appear. The most robust approach is to use only ASCII/LaTeX labels and avoid font dependencies entirely.

Workflow 2: Design (Lookup-Table Sizing)

Core Idea of the gm/ID Methodology

gm/ID (transconductance efficiency) is the pivot quantity linking circuit specifications to device dimensions.

Traditional design relies on long-channel model equations (gm = 2Id/Vov) to compute device sizes; errors are severe in advanced nodes (≤ 180 nm). The gm/ID method abandons equations in favor of simulation-generated design charts (lookup tables), using gm/ID as the single independent variable to express all key device performance parameters:

gm/ID  ──►  Id/W   (current-density curve → determines W)
       ──►  fT     (transit frequency → determines speed)
       ──►  gm·ro  (intrinsic gain → determines gain ceiling)
       ──►  Vgs    (uniquely determines the bias point)

All four curves are independent of transistor width W — devices of different W share exactly the same Id/W, fT, and gm·ro values at the same gm/ID point. This is the foundation of the methodology: the design chart needs to be generated only once (at unit width), and it applies directly to any W.

gm/ID is the trade-off axis for three design objectives:

Note: area and power optimization point in opposite directions — this is the central trade-off in gm/ID design.

Design Flow (Five Steps)

① Choose topology → ② Set L → ③ Choose gm/ID → ④ Derive gm / Id → ⑤ Look up Id/W → get W

① Choose topology Select the circuit structure (common-source, differential pair, cascode, common-gate, etc.) based on gain, bandwidth, and swing requirements.

② Set channel length L

• Need high speed (high fT) → choose minimum L
• Need high gain (high gm·ro) → increase L moderately (each doubling of L raises gm·ro by roughly 2–4×, with a corresponding fT reduction)
• Verify feasibility first: tbl.lookup('gmro', gmid_target) to check the upper bound; if unsatisfied, increase L before considering cascode or multi-stage topologies
• With a PMOS load, also verify PMOS ro: effective gain = gm_n × (ro_n ∥ ro_p)

③ Choose gm/ID Select the trade-off point on the fT–gm/ID and gm·ro–gm/ID curves based on the design priority:

tbl = GmIdTable('nmos45hp', W=1.0, L=0.045, vds=0.5)

# Check boundary values before deciding
print(f"fT   @ gmid=6  : {tbl.lookup('ft',   6.0)/1e9:.1f} GHz")
print(f"fT   @ gmid=15 : {tbl.lookup('ft',  15.0)/1e9:.1f} GHz")
print(f"gmro @ gmid=6  : {tbl.lookup('gmro',  6.0):.1f}")
print(f"gmro @ gmid=15 : {tbl.lookup('gmro', 15.0):.1f}")

④ Derive the required gm from circuit specs, then compute Id

# Example: BW and gain constraints → gm (ro-aware, see design example section)
Rout = 1 / (2 * 3.14159 * BW * CL)   # RL∥ro, set by bandwidth
gm   = Av / Rout                       # set by gain
Id   = gm / gmid                       # set by gm/ID

⑤ Look up Id/W to get W, round to 100 nm grid

id_w    = tbl.lookup('id_w', gmid)              # uA/um (W-independent)
W_exact = Id / (id_w * 1e-6)                   # um
W       = round(W_exact / 0.1) * 0.1           # round to 100 nm
# W     = math.ceil(W_exact / 0.1) * 0.1       # or round up (conservative margin)

After rounding, recompute Id = W × id_w and verify Av and BW; a ΔW < 5% deviation is usually negligible.

API Reference: GmIdTable

from design_gmoverid import GmIdTable, print_op

# First call: runs ngspice automatically and caches to logs/cache/ (JSON)
# Subsequent calls: reads from cache, ~0.05 s
tbl = GmIdTable('nmos180', W=10.0, L=0.18, vds=0.9)

Constructor parameters:

Single-quantity lookup (W-independent, depends only on gm/ID):

tbl.lookup('id_w', gmid)   # Id/W [uA/um]  <- use this to compute W
tbl.lookup('ft',   gmid)   # transit frequency fT [Hz]
tbl.lookup('gmro', gmid)   # intrinsic gain gm·ro
tbl.lookup('vgs',  gmid)   # Vgs (or |Vsg|) [V]  <- bias point
tbl.lookup('vov',  gmid)   # overdrive voltage Vov [V]
tbl.lookup('gm',   gmid)   # transconductance [S] (at reference width W)

Sizing: fix gm/ID + one constraint

op = tbl.size(gmid=15.0, Id=100e-6)   # given Id, solve for W
op = tbl.size(gmid=15.0, W=20.0)      # given W, solve for Id
op = tbl.size(gmid=15.0, gm=1.5e-3)   # given gm, solve for Id and W
print_op(op)

Constrained sizing: auto-search (highest gm/ID = lowest current)

op = tbl.size_from_ft(8e9,  W=20.0)    # fT >= 8 GHz, fixed W
op = tbl.size_from_gmro(35, Id=50e-6)  # gm·ro >= 35, fixed Id

• size_from_ft: suited for known-width, speed-constrained scenarios (LNA transconductor, OTA GBW)
• size_from_gmro: suited for high-gain stages (prefer fixing W to avoid excessively large W in weak inversion)

Keys returned by size():

model, L_um, W_um, Id_A, Vgs_V, Vov_V, gmid, gm_S, ft_Hz, gmro, id_w_Apm

Typical gm/ID Design Parameters by Node

nmos180 (Vds = 0.9 V, L = 180 nm)

nmos45hp / nmos22hp (HP, minimum L)

Vds has a minor effect on Id/W (advanced nodes have lower output impedance): ignore it during initial design, then fine-tune after simulation.

Design Example: 45 nm HP Common-Source Amplifier

Specifications (from textbook §5.2.1, 40 nm process, using PTM 45 nm HP as surrogate model):

1. Derive Design Constraints (ro-aware)

The gain equation includes the ro term and cannot be neglected:

|A_DC| = gm·(RL ∥ ro)

1/|A_DC| = 1/(gm·RL) + 1/(gm·ro)

=> gm·RL = 1 / (1/Av − 1/(gm·ro))

For 45 nm HP, intrinsic gain gm·ro ≈ 7 (confirmed via tbl.lookup('gmro', gmid)). Substituting:

gm·RL = 1 / (1/2 − 1/7) ≈ 2.80

Bandwidth determines the total output-node impedance (RL ∥ ro), i.e. the actual Rout:

Rout = RL ∥ ro = 1 / (2pi × BW × CL)
     = 1 / (2pi × 100 MHz × 10 pF) ≈ 159 Ohm

gm = Av / Rout = 2 / 159 ≈ 12.6 mA/V

RL = gm·RL / gm = 2.80 / 12.6 mS ≈ 222 Ohm

Why not simply set RL = Rout? Rout = 159 Ω is RL∥ro, not RL itself. RL must be larger than Rout and is back-calculated from the gain equation using the gm·ro term.

2. Size with GmIdTable

from design_gmoverid import GmIdTable
import math

VDD = 1.1;  Av = 2.0;  BW = 100e6;  CL = 10e-12

tbl = GmIdTable('nmos45hp', W=1.0, L=0.045, vds=0.5)

gmid = 10.0
gmro = tbl.lookup('gmro', gmid)   # => 7.11
id_w = tbl.lookup('id_w', gmid)   # => 155.4 uA/um
vgs  = tbl.lookup('vgs',  gmid)   # => 0.499 V  (bias Vgs)
vov  = tbl.lookup('vov',  gmid)   # => 0.244 V

# Step 1: derive gm and RL (ro-aware)
Rout   = 1 / (2 * 3.14159 * BW * CL)   # 159 Ohm = RL||ro
gm_RL  = 1 / (1/Av - 1/gmro)           # 2.80 (gain constraint including ro)
gm     = Av / Rout                      # 12.6 mA/V
RL     = gm_RL / gm                     # 222 Ohm

# Step 2: compute Id and W
Id      = gm / gmid                     # 1.26 mA
W_exact = Id / (id_w * 1e-6)           # 8.09 um
W       = round(W_exact / 0.1) * 0.1   # => 8.1 um (100 nm grid)

# Step 3: verify with rounded W
Id_r  = W * id_w * 1e-6                # 1.259 mA
gm_r  = Id_r * gmid                    # 12.59 mA/V
ro_r  = gmro / gm_r                    # 565 Ohm
Av_r  = gm_r * (RL * ro_r / (RL + ro_r))  # 2.00 ✓
Vd_DC = VDD - Id_r * RL                # 0.821 V  (margin 577 mV >> Vov 244 mV ✓)

3. Design Summary

4. Post-Design Simulation (hand off to ngspice skill)

With W, RL, and Vgs in hand, use the ngspice skill to build a .control block netlist and simulate:

• DC operating point: verify @m1[id], @m1[gm], @m1[gds] against design values
• AC frequency response: ac dec 200 1e5 1e10 → read vdb(vout) → verify low-frequency gain and −3 dB frequency
• Plotting: save frequency-gain data with wrdata, plot a Bode magnitude response using matplotlib

Self-Validation

Two complementary checks — run both after generating characterization plots.

Step 1 — Numerical self-check

python validate_gmoverid.py [model] [L_um]
python validate_gmoverid.py                    # nmos180 @ 180 nm
python validate_gmoverid.py nmos45hp 0.045

Five scalar tests, each prints PASS / FAIL:

For physics derivation and failure diagnosis see references/validation.md.

Step 2 — Visual inspection by LLM

Generate the four-quadrant plot, then show it to Claude and ask for a physics assessment. The expected trends for each panel are listed below — use them as the inspection checklist.

Panel (0,0) — gm/ID vs Vov

Red flags: peak > 35 (gm noise at low Id), plateau that never decays (missing strong-inversion region), kink at Vov=0 (Vth extraction error).

Note: for HP minimum-L nodes, DIBL raises the slope factor n toward 1.4–1.5, so the subthreshold ceiling legitimately sits closer to 25 V^-1 rather than 32 V^-1 — this is correct, not a model defect.

Panel (0,1) — Id/W vs gm/ID (log Y)

Red flags: any upward hook (fold-back from subthreshold rising side — fixed by _fall_mask), flat segment (resolution too coarse), curves widely separated by Vds (device not in saturation).

Panel (1,0) — fT vs gm/ID

Red flags: fT increases with gm/ID (sign inversion in gm or Cgg), non-monotone wiggles (ngspice convergence issue near threshold), fT > 1 THz or < 1 GHz at reasonable bias (Cgg extraction error).

Panel (1,1) — gm·ro vs gm/ID

Red flags: completely flat across all gm/ID (CLM ignored), gmro > 200 at moderate gm/ID for short-channel nodes (gds underestimated), kinks at specific Vgs points (sparse Vds-sweep interpolation artefact — use vgs_gds_results to enable finite-difference gds path).

Interpreting LLM visual feedback

LLM image analysis is powerful for shape and smoothness but requires physical context to interpret correctly:

• "Peak below BJT limit 38.6 V^-1" — expected; MOSFETs have n > 1. Correct ceiling is 1/(n·Ut) ≈ 25–32 V^-1.
• "Curves widely separated in panel (0,1) or (1,0)" — check Vds values; if all are well into saturation the separation should be < 20%.
• "gmro is very low (5–10)" — normal for HP minimum-L nodes; not a defect.
• "Curve does not reach weak inversion" — increase Vgs sweep range or lower id_thresh in sweep_vgs.

Reference Documents

See references/conventions.md for full details:

• §1–5 Project structure, symbol conventions, MODEL_INFO, netlist templates, data-dict keys
• §6–7 Sweep configuration constants (NODE_CFG), plotting conventions and figure list
• §8–9 Physical sanity-check values, common errors and fixes
• §10 Extending the skill (new nodes, new plot types, new channel lengths)
• §11 Full design API reference (GmIdTable, print_op, cache naming, unit conventions)