Analog Layout
analog-sizing picks W/L; this skill translates the sized schematic into a layout plan that actually survives process variation. Matching and noise immunity live or die on layout, not schematic.
When to use
- Current mirrors, differential pairs, bandgap references, DAC ladders
- Any circuit where σ(Vth) or σ(β) matching drives the spec
- Circuits near digital noise sources (guard rings / shielding needed)
- Before handing off to a layout engineer or Magic / Virtuoso
Inputs
- Sized schematic with device list (W, L, M, NF)
- Matching pairs / groups (which devices must match)
- Sensitivity list (which mismatches kill which spec)
- PDK rules (well spacing, latch-up, antenna)
- Floorplan block size budget
Workflow
- Identify matching groups — pairs, quads, ratios
- Choose matching style:
- Common-centroid 2×2, 4×4 for differential pairs
- Interdigitation ABAB or ABBA for current mirrors
- Centroid-preserving patterns for large ratios
- Dummy devices: at least 2 dummies on each side of a matching row
- Guard rings: n-well ring around PMOS, p-substrate ring around NMOS near digital
- Routing rules:
- Symmetric routing for differential signals
- Shielded routing for sensitive nodes (clock, bandgap output)
- Avoid crossing a matched row with unrelated metal
- Density / antenna: ensure fill rules and antenna rules met without breaking symmetry
- Post-layout extraction: plan for RC back-annotation → re-simulate → may need sizing iteration
Drawing a layout that actually closes A6 (measured recipe)
Everything above produces a plan. A6 (analog_a6_block_pv_check) does not grade plans: it
demands DRC == 0 AND a netgen LVS match, per block. eda_analog_layout cannot supply that —
it is honest that it returns status: SCAFFOLD because readspice + gds write places nothing.
So the layout is authored here, and three separate things have to be true. Each line below was
measured on sky130A (magic 8.3.679, device generator 1.0.599); the shape of the rule is
PDK-independent, the numbers are not.
SCOPE — this recipe is for the devices that do NOT have to match. Read it against the
Workflow above, not instead of it. §1 (one nf=1 device per schematic MOS) and §2 (each device's
columns inside its own slot) make a common-centroid quad and an interdigitated mirror
unconstructible: both of those patterns exist precisely to INTERLEAVE the devices this recipe
keeps apart. The two are not in competition — they apply to different devices:
| the devices whose mismatch drives a spec | every other device in the block |
|---|---|
| Workflow §1-§5 above: common-centroid 2×2 / 4×4, ABAB / ABBA interdigitation, ≥2 dummies per side, guard ring, symmetric routing | §1-§2 below: one nf=1 device per MOS, one slot each, generated two-layer routing |
A6 LVS will NOT match with dummies in the layout — the dummy is a device the schematic does not contain. Take the waiver route (/lvs-triage Batch waivers: "some analog dummies intentionally don't appear in schematic — document waiver") and record the waiver id in lvs_dummy_waiver. |
A6 LVS matches outright, which is what §1-§2 are optimised for |
State which one you took, in a field. A6's LVS compare is topology-only: it does not see a
centroid, does not see finger order, and dummies make it harder. So a block laid out as N
isolated devices closes A6 exactly as green as a fully matched one, and nothing in the tree tells
a reader which happened. Emit analog/<block>/layout_matching.json (schema under
Output format below). "matching_style": "none" is a legitimate, certifying answer — a level
shifter, a power switch and an ESD clamp have no matching group to build. What is not legitimate
is answering nothing: analog_a5_layout_check and analog_a6_block_pv_check both report the
class (declared_matched / declared_none / undisclosed) in their --json summary and on a
MATCHING: line, and A5 holds a disclosure that EXISTS to its own content
(A5_MATCHING_GROUP_DUMMIES_INSUFFICIENT is the ≥2-dummies rule above, executed;
A5_MATCHING_DUMMIES_LVS_UNRECONCILED is the waiver route, executed).
0. First check the netlist is layout-realizable at all
An A3 netlist containing an ideal V/I source, or an ideal R/C/L, inside the block
subcircuit can never pass A6, no matter how well it is drawn. The netgen PDK setup declares
device classes for the PDK's own res_* / cap_* subcircuits; it declares nothing that equates
an ideal SPICE primitive to a drawn device, so the compare cannot match. Measured: a block whose
netlist carried Vbias nbias VSS 0.8, R1 VOUT FB 100k, Cc nd2 VOUT 3p passed A3, A4 and A5
and made A6 unreachable by construction for three rounds.
This is a gate, not a review item. analog_a3_netlist_gen_check fails it as
A3_NETLIST_IDEAL_PRIMITIVE_IN_BLOCK — at A3, where the deck is written, not at A6 three steps
later where it reads as a layout defect. It is SPICE card grammar: V/I and the controlled and
behavioural sources E/F/G/H/B fire on the card letter alone; R/C/L fire only when
the value field holds a number or an expression, so R1 a b <pdk_res_model> w= l= and
XR1 a b <pdk_device> W= L= are both left alone. Cards at file scope and anything inside a
testbench subcircuit are out of scope — the testbench is where the fixes below SEND these
elements. The runner's own disclosed deterministic_stub deck is exempt.
Fix it upstream before drawing anything:
- ideal bias source → hoist to a block port the testbench drives (electrically identical);
- ideal R → the PDK poly/diff resistor with W/L solved for the same value;
- ideal C → the PDK MIM/VPP cap (arrayed if one instance is capped by the gencell's
lmax/wmax); - an ideal load element that the topology doc draws outside the block → move it to the testbench, do not delete it.
Then re-run A4 on the revised deck and quote the new corner numbers. A netlist change that
was never re-simulated is not a fix — and that is a gate too: a corner_results.json that
publishes netlist_sha256 / netlist_testbench_sha256 (as analog_real_corner_sweep does)
fails analog_a4_corner_sweep_check with A4_SWEEP_STALE_VS_NETLIST the moment either digest
stops recomputing from the file on disk. Every other subject-of-measurement rule at A4 is
answered once, when the artefact is written; this one is re-answered on every gate run.
1. One nf=1 device per schematic MOS, with W = W_schematic x M
(non-matched devices only — see SCOPE above. A matched group is interdigitated, and its finger count is set by the pattern, not by this rule.)
A multi-finger gencell exposes per-finger pins (D0_0, G0_0, S1…), and every one of them
extracts as a separate unconnected node — the single most common cause of
Final result: Top level cell failed pin matching. Setting conn_gates 1 merges the gates
only and leaves the source/drain comb unstrapped, so it does not rescue the compare.
An nf=1 gencell exposes exactly four clean ports: B D S G. And netgen expands m= into
parallel devices and merges them with parallel {w add} {l critical}, so a single wide finger is
LVS-equivalent to the schematic's multiplier. Verify that on your PDK in one command before
committing to it:
netgen -batch lvs "a.sp t" "b.sp t" <pdk>_setup.tcl out.log
# a.sp: X1 d g s b <nfet> W=4 L=1 m=4
# b.sp: X1 d g s b <nfet> W=16 L=1
# -> Final result: Circuits match uniquely.
For a very wide device, partition it into N parallel nf=1 devices whose widths sum to
W x M (12 x W=60 for W=6 m=120). Disclose the partition in
layout_matching.json's device_partitions[], not only in prose: it is LVS-equivalent and it is
a real connected structure, but its parasitics are not those of an interdigitated device.
analog_a5_layout_check does the arithmetic — A5_DEVICE_PARTITION_WIDTH_MISMATCH when the
declared layout widths do not sum to w_um x m, because netgen merges parallel devices by
ADDING their widths and so sees the sum, not the intent. A declared layout device with nf > 1
is recorded (multifinger_layout_devices) and not failed: whether a multi-finger gencell
extracts with per-finger pins is a property of one PDK's device generator, which is what the
one-command check above is for.
2. A routing discipline that is DRC-clean by construction
(non-matched devices only — see SCOPE above. "Each device's columns inside its own slot" is the opposite of interdigitation, and a matched row routes symmetrically instead.)
Do not hand-route. Use a two-layer discipline with a strict axis rule, so no two same-layer shapes of different nets ever share a coordinate band:
- met2 vertical only — one unique column per device terminal, at the terminal's own x;
- met3 horizontal only — one unique rail per net, above every device;
- via1 at the terminal, via2 at the (column, rail) crossing;
- devices spaced far enough apart that each device's columns stay inside its own slot.
This is generated from the netlist, not drawn, and it makes column/rail collisions a precondition check rather than a DRC round-trip.
3. The five PDK-specific facts that cost the most iterations
| symptom | cause | fix |
|---|---|---|
Via1 width < 52 / via2 width < 56 / via3 width < 0.32um |
magic's contact type carries the metal enclosure, so the drawn via must exceed cut + 2 x enclosure — not the cut size | draw via1/via2 >= 0.30 um, via3 >= 0.35 um |
Metal1 minimum area < 0.083um^2 on a short-L device |
the gencell's gate met1 strap is tiny at minimum L, at both top and bottom | enlarge both gate straps, not just the one carrying the port label |
| gate contact too close to the source/drain straps | the S/D met1 straps run the full device height and end just below the gate strap | raise the gate pad upward only; keep drc euclidean on so the diagonal clearance counts |
MiM cap spacing to unrelated metal3 < 1.34um |
the cap's top plate lives on met4 over its own met3 bottom plate | escape the top plate on met4 past the bottom plate's edge, then drop met4 -> met3 -> met2 |
Width of RPM/URPM < 1.27um on a bare resistor cell |
the narrowest poly-resistor gencell variant is DRC-dirty by itself | use the next wider variant and re-solve L for the same resistance |
Also: the substrate/well guard ring is the bulk terminal. Hang a short local-interconnect stub off the ring's own bar, put the li->met1 cut in the stub, and give it its own column — do not try to contact the ring under the device's met1 straps.
4. Prove it, then prove what it licensed
magic: drc euclidean on; drc style drc(full); drc check; drc catchup; drc list count total
magic: extract do local; extract all; ext2spice lvs; ext2spice -o <top>_layout.spice
netgen -batch lvs "<top>_layout.spice <top>" "<block>_lvs_schematic.spice <top>" <pdk>_setup.tcl
Copy netgen's own transcript verbatim into lvs.report and magic's own count into drc.report.
Then remember what A6 was holding back. A7 and A8 are PASS-VOIDED while A6 fails; the
moment A6 goes green they become live PASSes on whatever evidence happens to be sitting there.
Measured: a pre_vs_post.json whose every delta_pct was 0.0 because the post values had been
copied from the pre values, and a hardmacro LEF/GDS streamed from the superseded unrouted layout.
Rebuild both from the routed layout:
- A7 —
extract all+ext2spice cthresh 0+rthresh infinitegives a real post-layout netlist; simulate it with the same stimulus and the same.measstatements as A4. If every delta is exactly zero, the numbers were copied — and both gates overpre_vs_post.jsonnow refuse it: a comparison whose every compared spec's post value equals its pre value certifies only if it NAMES the post-layout artefact its post column was simulated from and that artefact resolves on disk (A7_POSTSIM_ALL_ZERO_DELTA_UNEVIDENCEDinanalog_a7_post_layout_resim_check,PRE_VS_POST_ALL_ZERO_DELTA_UNEVIDENCEDinanalog_pre_vs_post_layout_check). Name the extraction; a0.0column with nothing behind it is the one input that used to score better than every honest comparison. - A8 —
gds write+lef writefrom the routed cell. Two traps: magic'slef writeframe is the magic cell bbox, which can include magic-only layers (sky130pwell) that stream no GDS, so the LEF and GDS land in different frames — set an explicitFIXED_BBOXequal to the GDS extent first. And magic'slef writeemits noUSE POWER/USE GROUND: re-attachDIRECTIONandUSEto every pin. This one is also gated now — an abstract that types NO pin at all is no longer read as an abstract with nothing to declare, andl21_macro_supply_rail_declared_checkrecovers the typing from the macro's own Libertypg_pin/pg_typeand fromL21.power_domains[]before reporting the real rail finding, or says the contract is UNVERIFIABLE (rule L21-5, waiver keyl21_macro_lef_pin_use_absent_disclosed). An untyped abstract used to SKIP with a message byte-identical to a design that has no macro at all.
Still unowned here, named rather than dropped. Two of §4's claims are not yet programs, and neither is silently assumed:
- A7/A8 staleness against the routed layout. The rules above catch a COPIED post column and
an UNTYPED abstract; neither catches evidence that was honestly measured against a
superseded layout. A4 has the equivalent rule (
A4_SWEEP_STALE_VS_NETLIST) only because its producer stamps a digest of its input; no A7 or A8 producer records the layout digest it was built from, so the check has nothing to recompute against. It needs the producer change first — do not substitute mtime, which a fresh checkout rewrites. - The
netgen -batch lvsparallel-merge self-test in §1. It is a real probe and it belongs in a program, but it needs netgen and a PDK setup TCL inside the EDA container. A program that cannot run its tool would either SKIP always — certifying nothing while looking like a gate — or assert the answer it was meant to measure. Run the command by hand until it is wired to a container-backed tool call.
Output format
layout/<block>_layout_plan.md:- Device placement diagram (ASCII or SVG)
- Matching pattern per group
- Routing rules
- Dummy / guard-ring list
- Known risks
analog/<block>/layout_matching.json— the same three facts as a FIELD, because A6's LVS compare is topology-only and cannot tell a matched layout from an unmatched one. Read byanalog_a5_layout_check(rules) andanalog_a6_block_pv_check(record). Every rule fires only on a file that exists, so writing one costs nothing except having to mean it:{ "block": "<block>", "matching_style": "common_centroid" | "interdigitated" | "none", "matched_groups": [ {"name": "input_pair", "devices": ["Mn1", "Mn2"], "style": "common_centroid", "dummies_per_side": 2} ], "lvs_dummy_waiver": "<ticket>", "device_partitions": [ {"schematic_device": "Mpass", "w_um": 6.0, "m": 120, "layout_devices": [{"w_um": 60.0, "nf": 1}]} ], "note": "free text" }field rule matching_styleabsent / emptyA5_MATCHING_DISCLOSURE_MALFORMED"none"with groups listed, or a style with no groupA5_MATCHING_STYLE_GROUPS_CONTRADICTa matched group with dummies_per_side< 2, or none declaredA5_MATCHING_GROUP_DUMMIES_INSUFFICIENTdummies declared, no lvs_dummy_waiverA5_MATCHING_DUMMIES_LVS_UNRECONCILEDlayout_devices[].w_umdoes not sum tow_um x mA5_DEVICE_PARTITION_WIDTH_MISMATCHomitted entirely no rule — undisclosed, reported in both gates' summary and on theMATCHING:line
Technical basis
Classic references: Razavi "Design of Analog CMOS Integrated Circuits" ch. on layout, Hastings "The Art of Analog Layout", Pelgrom matching model. Process-specific matching coefficients come from the PDK.
Handoff
- Sized devices → input to this skill came from
/analog-sizing - LVS verification →
/lvs-triage - DRC verification →
/drc-fix - Post-layout resim →
/ams-sim
Captured by benchmark-enhancement-capture — 2026-09-01 (u_hawaii_adc A5 convergence: two blocks to netgen match-unique + KLayout DRC 0 on a native PDK)
The deterministic half of every rule below ships in
programs/magic_gencell_layout_lib.py (LAW numbers reference the campaign's
A5_STATUS.md in benchmark-data uhadc/a5-layout-generator, where the full
24-law evidence trail and the working generator live). Use the library; do
not re-derive the parsing or the geometry checks by hand.
Skill: Magic .mag coordinate spaces — read magscale PER FILE, rlabels are always internal (LAWS #1/#22)
Pattern: use transforms and rect lines are in lambda ONLY when the
file has no magscale header; magic silently writes magscale 1 2 the
moment any geometry sits off the lambda grid (a half-lambda column of an
odd-length device is enough), and then every transform/rect in THAT file is
2x. Gencell children split by family (measured: MOS/resistor children carried
the header, MIM-cap children did not). rlabel coordinates are internal
(2x) in BOTH kinds of file.
When to apply: any scripted Magic layout — before trusting a single parsed coordinate.
What to do: parse with magic_gencell_layout_lib.mag_scale /
parse_use_transforms / parse_rlabels. Never assume a family of .mag
files shares a unit.
Worked failure: reading a magscale 1 2 parent as lambda doubled every
origin — the entire wiring layer painted off the devices; extraction
returned one single-pin net per terminal (51 nets, zero joins) with NO tool
error anywhere.
Why this is GENERAL: the header semantics are magic's own file format, independent of PDK; any tech whose gencell emits off-grid geometry trips it.
Skill: exit-ladder discipline — stagger tied taps, allocate descents ascending AND obstacle-aware (LAWS #9/#23)
Pattern: a right-side exit ladder is short-proof only while rung y is
strictly ascending with rank ("a rung at Y_i can only cross a descent whose
top is below Y_i"). Two facts break a naive implementation: (a) D and S
labels of one MOS share a y, so two rungs go co-linear and short; (b) a
descent placed blindly at origin + pitch*(rank-1) ignores everything
between its tap and its rail — including other rows' tap pads.
When to apply: any tight-pitch cell whose contact columns cannot host per-column vertical lanes (measured threshold: column pitch below wire width + spacing + via pad).
What to do: stagger_ladder_taps (ties raised >= wire+space along the
column's own full-height M2 — LAW #9 guarantees the column carries it),
then allocate_descent per rank with every terminal and ladder point of
OTHER nets as obstacles. Afterwards run cross_net_overlaps on the wiring
manifest — an empty result is the necessary condition, not a formality.
Worked failure: 28 same-device cross-net shorts from co-linear rungs, plus one macro's blind descent through a neighbour row's tap pads; "Top level cell failed pin matching" was the only symptom LVS showed.
Why this is GENERAL: the proof obligation (ascending y, foreign-point clearance) is geometry, not PDK; only the pitch numbers come from the deck and they are parameters.
Skill: MIM-cap plate exits derive from the MEASURED plate bbox, never fixed offsets (LAW #24)
Pattern: the top-plate via stack auto-paints a pad on the bottom-plate metal at its lowest level, so its lateral exit must clear the cap's own bottom plate; the bottom-plate exit must clear the NEIGHBOUR's plate. Fixed offsets are chip-luck: they encode one cap size.
When to apply: wiring any MIM/MOM cap whose plates are on routing metals; sizing changed by even one micron re-opens this.
What to do: cap_plate_exits(plate_bbox) with the deck's clearances;
verify with cross_net_overlaps including the via3+/metal5 links.
Worked failure: three plate welds (two caps internally shorted, one
onto its neighbour) — found only by a flattened-layout island probe,
because l2n.shapes_of_net returns CHILD-LOCAL coordinates and an
unflattened weld analysis is garbage.
Why this is GENERAL: the auto-pad behaviour is the via-stack's, the clearance is a deck number, and the bbox is measured per instance.
Skill: magic extraction is NOT connectivity-authoritative on gencell PDKs — sign off with the PDK's own LVS, quantize the comparison to the grid
Pattern: magic's extraction (hierarchical AND flat) does not reliably
merge parent-painted metal with gencell contact stacks (measured: 93
isolated terminal nets where KLayout found 9). Sign-off connectivity is the
PDK's KLayout LVS + netgen. At the comparison, the drawn mask is on the
manufacturing grid while derived netlist parameters may not be; the PDK
netgen deck declares a tolerance for this but the clause is inert in
current netgen builds (control-run measured). Quantize the COMPARISON copy
with grid_snap_spice_params; never touch the design netlist.
When to apply: every A5/A6 LVS on a Magic-gencell layout.
why_not_bucket_a (the residual): WHICH netlist transform closes a given mismatch (canonical net renaming, X-to-primitive conversion, parameter bookkeeping strips) depends on the PDK's extraction dialect and stays a per-PDK recorded recipe; the library ships the invariant pieces only.
Captured by benchmark-enhancement-capture 2026-09-01 (u_hawaii_adc).