Analog Topology Select
Given a block's spec.json (from analog-spec-extract) and PDK device characteristics, recommends a circuit topology. This is the highest-leverage decision in analog design — wrong topology wastes 10+ sizing iterations; right topology converges in 1-2.
When to use
- Step A2 of the analog track
- After
analog-spec-extracthas produced spec.json per block - When the user asks "what circuit should I use for this LDO?"
Inputs
analog/<block>/spec.json— fromanalog-spec-extract- PDK device characteristics — do NOT retype Vth/supply constants here.
They are a single deterministic source-of-truth in
programs/pdk_registry.jsonunder the per-PDKanalog_device_paramsblock (vth_n_v,vth_p_v,nominal_supply_v). Look them up by PDK name; the A2 gate (programs/analog_a2_topology_select_check.py) andanalog-sizing-loopconsume the same registry field. (canonical reference: GF180 nfet≈0.65V/pfet≈0.70V/3.3V; SKY130 nfet≈0.45V/pfet≈0.47V/1.8V.) 2b. The MEASURED half of that same field (vibe-ic#1962) —analog_device_params.measured, read withprograms/pdk_analog_device_params.pyand produced byprograms/pdk_analog_characterize.py. It carries what the declared constants above do not:k_prime_n_ua_per_v2/k_prime_p_ua_per_v2, the threshold the square-law fit implies,vgs_at_id_*_v,rsheet_ohm_per_sq(+r_end_ohm) andcap_area_ff_per_um2(+cap_perim_ff_per_um), per process corner, each attributed to the primitive and the model-lib section it was measured on and each carrying the fit residual that says how well the model describes the device.analog_a2_topology_emitputs these intotopology.jsonaspdk_measured_params, so a librarydevice_param_exprsentry may be written against one of them by name. A family with NO record quotes nothing and says so — never a default, never a neighbouring family's number. - Constraints: power budget, area budget, accuracy requirements
Proven topologies (GF180, verified via SPICE)
These templates are from analog-sizing/PRACTICAL_NOTES.md — all verified working on GF180 with ngspice:
LDO
Recommended: NMOS-input diff pair + PMOS active mirror + PMOS series pass transistor
- Why NMOS input (not PMOS): at Vth=0.65V, PMOS input pair common-mode range is too narrow for 1.8V output feedback
- Feedback: resistor divider (R1, R2) from Vout to inverting input
- Compensation: Miller Cc from output to diff-pair drain
- Anti-pattern: PMOS-input diff pair — fails at GF180 Vth levels
Oscillator
Recommended: Current-starved 5-stage ring oscillator
- Frequency set by bias current (Ibias) → tunable via current mirror ratio
- No inductors needed (GF180 has no spiral inductor models)
- Anti-pattern: LC oscillator — no inductor in GF180
POR (Power-On Reset)
Recommended: PMOS diode offset + resistor divider + 2-inverter chain + weak PMOS feedback + RC delay
- PMOS diode provides fixed |Vtp| offset (~0.7V) → trip point independent of VDD
- Hysteresis via weak PMOS feedback from output to inverter chain
- RC delay for clean release timing
- Anti-pattern: Schmitt trigger — inverter threshold tracks VDD proportionally, making trip point unreliable
Bandgap Reference
Recommended: Brokaw bandgap (CTAT + PTAT summing)
- Two BJTs (or parasitic PNP in GF180) with different current densities
- CTAT from Vbe, PTAT from ΔVbe × R ratio
- Output: ~1.2V reference (process-independent)
Comparator
Recommended: Folded-cascode + latch output stage
- Folded-cascode for high gain + wide input range
- Latch for fast decision with digital-compatible output
Workflow
- Load
spec.jsonand identify block type - Map to 2-3 candidate topologies from the library above
- Evaluate each against:
- Supply voltage headroom. Vth and the nominal supply are deterministic
PDK constants — read
vth_n_v/vth_p_v/nominal_supply_vfromprograms/pdk_registry.json(analog_device_params), do NOT guess. When the family has been characterized, prefer the MEASUREDvth_*_extracted_vfor a headroom argument about a device in saturation: it is the threshold the square-law fit of THIS process implies at a stated bias, whereas the declared value is a model-card constant. Both are in the record; quote which one you used. The headroom feasibility itself stays judgment: you must read each candidate's schematic to count how many devices are stacked between the rails and pick realistic Vdsat/Vov margins (the spec gives non_stackedorVdsatfield, so there is no deterministic value to program — see "Do not / headroom" below). Then checkavailable = Vdd - Σ(Vth + Vdsat over the stack) ≥ 0per candidate. - Power budget vs. required performance
- Area constraints
- PDK device availability (e.g., no inductors in GF180)
- Supply voltage headroom. Vth and the nominal supply are deterministic
PDK constants — read
- Select best topology with justification
- Output
topology.mdwith:- Schematic description (ASCII art or text)
- Key device roles
- Trade-off analysis
- Why alternatives were rejected
- Flag if no known topology fits → escalation
Output format
analog/<block>/topology.md — one file per block:
# Topology — <block_name>
## Selected: <topology_name>
## Schematic
<ASCII schematic or description>
## Device roles
| Device | Role | Key constraint |
|--------|------|----------------|
| M1,M2 | Input diff pair | Vgs > Vth + Vdsat_tail |
| ... | | |
## Trade-off analysis
| Candidate | Pros | Cons | Verdict |
|-----------|------|------|---------|
| NMOS-input OTA | Wide CM range at 3.3V | Higher noise | Selected |
| PMOS-input OTA | Lower noise | CM range too narrow | Rejected |
## PDK constraints applied
- Vth(N), Vth(P), Vdd — quote the exact values you looked up from
pdk_registry.json analog_device_params for this PDK → requires ...
- No inductors → ring oscillator instead of LC
Do not
- Do not skip the trade-off analysis — topology selection without justification is the #1 cause of wasted iterations
- Do not assume textbook Vth (0.3-0.5V) — read the real per-PDK Vth from
programs/pdk_registry.json(analog_device_params); GF180's ≈0.65V invalidates many textbook topologies - Do not recommend topologies that require devices not in the PDK
- headroom: Do not invent a fixed
n_stackedor Vdsat margin to make the headroom check a program. The Vth/supply constants are programmatic (registry), but how many devices stack between the rails and the per-device Vdsat depend on the candidate schematic and bias point — that is genuine analog judgment, not a deterministic table lookup.
Handoff
analog/<block>/topology.md→/analog-sizing(Step A2, continued)- Rejected topologies documented for future reference
Captured rule
Keep gain, noise and current allocations in the topology as design data, with
their formulas and document sources. They are not acceptance limits. A native
budget_studies declaration names the document row and the measured scope;
compare supply usage with that row's maximum current, never with its minimum
load or a private allocation. Explicitly retained native results may be reused
when the circuit and conditions have not changed; read the simulator log for
the number. Otherwise run the declared analysis on the delivered circuit:
OP for a static bias target, transient averaging for clocked operating current.
For an ENOB bound, use SINAD = 6.02 * ENOB + 1.76 dB and the named documented full-scale amplitude, sampling rate and OSR to derive the allowed in-band noise. Compare total input-referred RMS noise from the simulator's noise analysis over that band. Missing document inputs or native noise remain NOT_MEASURED; do not substitute static power, a single sub-block, or a transient ENOB result. No documented gain bound means no gain check. The later ENOB requirement stays unchanged and still needs its own native measurement.
Captured by benchmark-enhancement-capture — 2026-05-28 (RTLLM Shape B + benchmark_clean + CVDP cross-step capture)
Skill: ΔΣ modulator topology — 2nd-order SC CIFB with 1-bit quantizer + feedback DAC
Pattern: When the analog block class is delta_sigma_adc or delta_sigma_modulator, the canonical SC topology that Analog A2's primitive panel must recognize is: TWO switched-capacitor integrators (NMOS-input two-stage Miller OTA + Cs/Ci + two-phase non-overlapping switches) + 1-bit clocked comparator (preamp + cross-coupled regenerative latch) + 1-bit feedback DAC (±Vref select switches). This is the CIFB (cascade-of-integrators with distributed feedback) topology, the textbook 2nd-order ΔΣ.
When to apply: Analog A2 topology selection for blocks whose L5 spec names delta_sigma, delta-sigma, incremental ΔΣ, or SC modulator. Also applicable for sub-blocks of a higher-order ΔΣ ADC chain.
What to do: Author topology.md listing: opamp (two-stage Miller, NMOS input pair, PMOS load + class-AB output), comparator (StrongARM-style or preamp+latch), SC integrator (Cs/Ci ratio sized for stability + gain), 1-bit DAC (CMOS analog mux on Vref±). For higher orders, extend to cascaded integrators with distributed/feedforward coefficients.
Worked example (from u_hawaii_adc): u_hawaii_adc A2: prior A2 panel rejected delta_sigma block-type (no transistor/primitive keyword match). The skill-authored topology.md named opamp/comparator/SC integrator/DAC as transistor-level primitives → A2 PASS. Real ngspice on the resulting netlist showed UGBW 1.81-14.93 MHz across 9 corners (> 1 MHz fclk → integrator settles).
Why this is GENERAL: Standard textbook ΔΣ topology (any Schreier / Norsworthy reference). Applies across audio ADCs, sensor readout, incremental high-resolution conversion. Doesn't depend on the specific spec target.
Captured by benchmark-enhancement-capture 2026-05-28.
Captured by benchmark-enhancement-capture — 2026-06-06 (analog block-list sanity vs L5 enumeration)
Skill: spurious-block sanity check — confirm against the L5 type enumeration BEFORE sizing
Pattern: When L5 explicitly states the number of analog block TYPES (a sentence like "the chip contains N analog blocks", or a "Block A / Block B / …" table whose header rows enumerate the blocks), treat that enumeration as the authoritative block roster. Any candidate block carried into A2 whose:
specisnull(no electrical numbers were ever bound to it in A1), AND- only evidence is a product-name keyword match from L1 (the block "exists" solely because a token in the datasheet/product title looked like an analog class),
is presumed SPURIOUS. Do NOT spend any sizing/topology compute on it until you have confirmed it against the L5 enumeration. If the L5 type list does not include a block of that class, drop the candidate (it was an L1-name false-positive); if L5 does enumerate it, the block is real — proceed, and bind its spec from the L5 entry.
Multiplicity rule: a block's multiplicity (×N — "four identical comparators",
"dual LDO") must come from that block's OWN table-header / enumeration entry, never
from a sibling block's evidence paragraph. A "×4" stated in the LDO row does not make the
comparator a ×4 block. Read each block's own enumeration row for its own count; if the
block has no count of its own, it is ×1.
When to apply: A2 topology selection, as the first sanity pass over the A1 block list, before mapping any candidate to candidate topologies. Cheapest possible gate — one read of the L5 enumeration saves 10+ wasted sizing iterations on a phantom block.
What to do:
- Read the L5 block-type enumeration (count sentence and/or Block A/B table headers).
- For each A1 candidate with
spec == null, check whether its only provenance is an L1 product-name keyword. If so, require an L5 enumeration entry to keep it. - Drop candidates with no L5 backing; keep + spec-bind those with an L5 entry.
- Read each kept block's OWN enumeration row for its multiplicity; default ×1.
Why this is GENERAL: every analog datasheet that enumerates its blocks (count or Block-table headers) admits this check; it names no chip and depends on no SKU. The provenance test (null-spec + name-only evidence) is a structural property of the A1 output, not a hard-coded block list.
why_not_bucket_a (cannot be a deterministic deny-list): whether a product-name-matched
block is real cannot be safely deny-listed. Some chips' blocks genuinely appear only in the
datasheet/product name (the name is the evidence), so a static "ignore name-only blocks"
rule would silently delete real blocks. The decision requires reading the L5 enumeration
to confirm-or-deny each candidate — a judgment over prose/table semantics, not a regex over
the block name. The program's job stays "surface every name-keyword hit" (recall floor in
analog-spec-extract); the confirm-against-L5 judgment is the residual LLM step.
Captured by benchmark-enhancement-capture 2026-06-06 (#466B).
Captured rule — a clock-stopped OP is not operating supply current
Before comparing a clocked block's supply usage with its document bound, read which analysis produced the number. A static operating point with the clock held low does not measure average current during conversion. Digital nodes can also settle at a metastable DC root. Retain that OP as a bias diagnostic; do not use it to decide a clocked-operation current requirement.
Use the delivered circuit and its declared operating stimulus. Read the native transient average over the declared steady conversion window, spanning whole clock cycles after startup. Compare only the corners and operating conditions actually measured, against the unchanged document limit. Missing operating measurements remain NOT_MEASURED.
On the delta-sigma ADC at 1.3 V, clock-stopped FF/-40 C OP reported 1.01247 mA against the 1 mA supply bound; FF/27 C reached a different DC root at 3.65155 mA. Neither is a conversion-current measurement. A completed TT/125 C transient at the declared 1.2823 MHz maximum clock reported 0.581586 mA over 400.062 to 799.345 us. That result is one operating corner, not an FF comparison or a full-corner PASS. The clocked-current producer and remaining corner results belong to the analog implementation recovery.
Captured rule — an aborted transient's printed zero is not a measurement
Read the native transient log and the measurement's actual interval before
using its value. ngspice can exit 0, print a numeric zero and even print its
normal completion banner after Timestep too small and tran simulation(s) aborted. An average whose reported end does not reach the declared window
has not measured that window. Keep the abort and partial trace as diagnostics;
report the operating quantity as NOT_MEASURED, not as zero current or a
passing budget. This does not change a simulator option or the original bound.
The ADC FF/-40 C, 1.3 V transient stopped at 68 us. Its native log reported
supply_current = 0.00000e+00 from=5.13000e-04 to=6.80000e-05 and
ngspice-47 done, while the command returned 0. The requested average began
at 513 us, after the analysis had already stopped. The zero therefore cannot
be compared with the 1 mA limit. Diagnose the convergence failure separately;
only a completed native interval can supply the operating measurement.