Analog Sizing
Given an analog topology (schematic or netlist) and a performance spec (gain, UGB, phase margin, noise, power, area), propose transistor W/L sizes and bias currents that meet the spec. Acts as the LLM copilot layer over classical analog sizing tools and SPICE.
When to use
Trigger when the user:
- Has a schematic (op-amp, LDO, bandgap, comparator, PLL CP) and needs sizes
- Wants to explore a design space before running SPICE sweeps
- Asks for gm/Id-based sizing intuition
- Needs a starting point for an optimizer
Inputs to gather
- Topology (schematic, netlist, or textual description)
- Process and models (e.g., TSMC 180nm, 65nm, FinFET node)
- Spec table: gain (dB), UGB (Hz), PM (°), noise, power budget, supply, load
- Starting point or "greenfield"
- Constraint: minimize power, minimize area, or balanced
The process constants are MEASURED, not re-derived here (vibe-ic#1962)
Do not derive k', the threshold, the resistor sheet or the MiM density by
hand, and do not quote them from memory. They are a deterministic measurement
of the target PDK's own models, taken once and published with their provenance:
# read what has already been measured for the target family
python3 plugins/vibe-ic/programs/pdk_analog_device_params.py --pdk <family>
# measure (or refresh) a family that carries no record yet
python3 plugins/vibe-ic/programs/pdk_analog_characterize.py \
--pdk <family> --container vibeic-eda --corners typ,slow,fast
The record lives under each PDK entry's analog_device_params.measured in
programs/pdk_registry.json — or, for a PDK STAGED INTO the design (which may
not be publishable), under analog/_pdk_char/analog_device_params.json in the
project, which outranks the shipped record. What it gives you, per corner:
| constant |
what it is |
k_prime_n_ua_per_v2 / k_prime_p_ua_per_v2 |
µ·Cox, from a two-point square-law fit in saturation |
vth_n_extracted_v / vth_p_extracted_v |
the threshold THAT fit implies, which is not the model card's Vth0 |
vgs_at_id_n_v / vgs_at_id_p_v |
the gate drive a stated current buys in a stated geometry — the bias-chain handle |
rsheet_ohm_per_sq, r_end_ohm, r_per_um_ohm |
the resistor sheet, and the end/contact term a single-device measurement hides in it |
cap_area_ff_per_um2, cap_perim_ff_per_um |
the capacitance density, and the fringe a single-plate measurement hides in it |
Read the fit residual before you use a k'. It is the relative error of an
interior bias point the fit did not use. A residual of a few percent means the
square law describes the device; a large one means the value is the best
square-law fit of a device that is not square-law, and a gm/Id argument is the
better tool at that operating point.
Read not_measured too. A constant that is absent is absent because the
measurement refused, and the reason is stated — most commonly a resistor
flavour that ignores the drawn width, for which r_per_um_ohm is published
instead of a sheet. Never substitute a neighbouring family's number for one
this family does not carry.
What stays YOUR judgment, and is not in that record: which spec binds which
device, how to split the current budget, what overdrive to spend, which corner
the block must close at, and whether the topology can meet the spec at all.
The measurement removes the arithmetic, not the design.
Sizing workflow
- Identify the signal path and loading — what sets gain, what sets BW, what dominates noise
- Pick operating regions — strong inversion for speed, weak/moderate for efficiency
- Use gm/Id methodology — pick a gm/Id target per device, derive Id from gm, then W/L
- Size mirrors and biases — match ratios, ensure headroom
- Compensation — for op-amps, pick Cc and nulling resistor for target PM
- Sanity-check — walk through the spec line by line against the proposed sizes
- Recommend SPICE verification — always
Output format
# Analog Sizing — <topology>
Process: <node>
Supply: <V>
## Operating point plan
| Device | Role | gm/Id | Id (µA) | W (µm) | L (µm) | Region |
|--------|------------|-------|---------|--------|--------|--------|
| M1 | input pair | 15 | 20 | 40 | 0.5 | moderate |
| ... |
## Predicted performance (gm/Id estimates)
| Spec | Target | Estimate | Margin |
|------|--------|----------|--------|
| Gain | 60 dB | ~62 dB | + |
| UGB | 10 MHz | ~11 MHz | + |
| PM | 60° | ~62° | + |
| Power | 200 µW | ~180 µW | + |
## Compensation
Cc = ..., Rz = ...
## Verification
Run SPICE: <suggested testbench(es) — AC, transient, noise, corners>
Technical basis
Grounded in ADO-LLM, Maieutic Semiconductor's analog copilot direction, and gm/Id methodology (Silveira, Flandre, Jespers). Core idea: analog sizing is a constrained optimization where good initial conditions matter more than the optimizer — and LLMs trained on analog literature are very good at initial conditions.
Do not
- Do not claim final sign-off from hand analysis; always recommend SPICE
- Do not ignore corners — flag typical/SS/FF/hot/cold as an explicit next step
- Do not propose sizes that violate matching, mirror ratios, or common-mode constraints
1---2name: analog-sizing3description: Size transistors in an analog circuit topology to meet performance specs (gain, bandwidth, noise, power). Use when the user says "size this amplifier", "analog sizing", "op-amp design", "bias point", "find W/L", or shares a schematic and a spec table.4metadata:5 source: https://skillmd.com/skills/vibeic/analog-sizing6---78# Analog Sizing910Given an analog topology (schematic or netlist) and a performance spec (gain, UGB, phase margin, noise, power, area), propose transistor W/L sizes and bias currents that meet the spec. Acts as the LLM copilot layer over classical analog sizing tools and SPICE.1112## When to use1314Trigger when the user:15- Has a schematic (op-amp, LDO, bandgap, comparator, PLL CP) and needs sizes16- Wants to explore a design space before running SPICE sweeps17- Asks for gm/Id-based sizing intuition18- Needs a starting point for an optimizer1920## Inputs to gather21221. Topology (schematic, netlist, or textual description)232. Process and models (e.g., TSMC 180nm, 65nm, FinFET node)243. Spec table: gain (dB), UGB (Hz), PM (°), noise, power budget, supply, load254. Starting point or "greenfield"265. Constraint: minimize power, minimize area, or balanced2728## The process constants are MEASURED, not re-derived here (vibe-ic#1962)2930Do **not** derive k', the threshold, the resistor sheet or the MiM density by31hand, and do not quote them from memory. They are a deterministic measurement32of the target PDK's own models, taken once and published with their provenance:3334```bash35# read what has already been measured for the target family36python3 plugins/vibe-ic/programs/pdk_analog_device_params.py --pdk <family>3738# measure (or refresh) a family that carries no record yet39python3 plugins/vibe-ic/programs/pdk_analog_characterize.py \40 --pdk <family> --container vibeic-eda --corners typ,slow,fast41```4243The record lives under each PDK entry's `analog_device_params.measured` in44`programs/pdk_registry.json` — or, for a PDK STAGED INTO the design (which may45not be publishable), under `analog/_pdk_char/analog_device_params.json` in the46project, which outranks the shipped record. What it gives you, per corner:4748| constant | what it is |49|---|---|50| `k_prime_n_ua_per_v2` / `k_prime_p_ua_per_v2` | µ·Cox, from a two-point square-law fit in saturation |51| `vth_n_extracted_v` / `vth_p_extracted_v` | the threshold THAT fit implies, which is not the model card's Vth0 |52| `vgs_at_id_n_v` / `vgs_at_id_p_v` | the gate drive a stated current buys in a stated geometry — the bias-chain handle |53| `rsheet_ohm_per_sq`, `r_end_ohm`, `r_per_um_ohm` | the resistor sheet, and the end/contact term a single-device measurement hides in it |54| `cap_area_ff_per_um2`, `cap_perim_ff_per_um` | the capacitance density, and the fringe a single-plate measurement hides in it |5556**Read the `fit` residual before you use a k'.** It is the relative error of an57interior bias point the fit did not use. A residual of a few percent means the58square law describes the device; a large one means the value is the best59square-law fit of a device that is not square-law, and a gm/Id argument is the60better tool at that operating point.6162**Read `not_measured` too.** A constant that is absent is absent because the63measurement refused, and the reason is stated — most commonly a resistor64flavour that ignores the drawn width, for which `r_per_um_ohm` is published65instead of a sheet. Never substitute a neighbouring family's number for one66this family does not carry.6768What stays YOUR judgment, and is not in that record: which spec binds which69device, how to split the current budget, what overdrive to spend, which corner70the block must close at, and whether the topology can meet the spec at all.71The measurement removes the arithmetic, not the design.7273## Sizing workflow74751. **Identify the signal path and loading** — what sets gain, what sets BW, what dominates noise762. **Pick operating regions** — strong inversion for speed, weak/moderate for efficiency773. **Use gm/Id methodology** — pick a gm/Id target per device, derive Id from gm, then W/L784. **Size mirrors and biases** — match ratios, ensure headroom795. **Compensation** — for op-amps, pick Cc and nulling resistor for target PM806. **Sanity-check** — walk through the spec line by line against the proposed sizes817. **Recommend SPICE verification** — always8283## Output format8485```86# Analog Sizing — <topology>8788Process: <node>89Supply: <V>9091## Operating point plan92| Device | Role | gm/Id | Id (µA) | W (µm) | L (µm) | Region |93|--------|------------|-------|---------|--------|--------|--------|94| M1 | input pair | 15 | 20 | 40 | 0.5 | moderate |95| ... |9697## Predicted performance (gm/Id estimates)98| Spec | Target | Estimate | Margin |99|------|--------|----------|--------|100| Gain | 60 dB | ~62 dB | + |101| UGB | 10 MHz | ~11 MHz | + |102| PM | 60° | ~62° | + |103| Power | 200 µW | ~180 µW | + |104105## Compensation106Cc = ..., Rz = ...107108## Verification109Run SPICE: <suggested testbench(es) — AC, transient, noise, corners>110```111112## Technical basis113114Grounded in ADO-LLM, Maieutic Semiconductor's analog copilot direction, and gm/Id methodology (Silveira, Flandre, Jespers). Core idea: analog sizing is a constrained optimization where good initial conditions matter more than the optimizer — and LLMs trained on analog literature are very good at initial conditions.115116## Do not117118- Do not claim final sign-off from hand analysis; always recommend SPICE119- Do not ignore corners — flag typical/SS/FF/hot/cold as an explicit next step120- Do not propose sizes that violate matching, mirror ratios, or common-mode constraints