Skip to content

analog M4b-psp: the StrongARM comparator on PSP103 tables at 50 C against ngspice -- sign, times, flip point, imbalance sensitivity (Closes #10856) - #10884

Merged
gHashTag merged 2 commits into
masterfrom
claude/analog-m4b-psp-10856
Oct 19, 2026
Merged

analog M4b-psp: the StrongARM comparator on PSP103 tables at 50 C against ngspice -- sign, times, flip point, imbalance sensitivity (Closes #10856)#10884
gHashTag merged 2 commits into
masterfrom
claude/analog-m4b-psp-10856

Conversation

@gHashTag

@gHashTag gHashTag commented Oct 19, 2026

Copy link
Copy Markdown
Owner

Closes #10856. Part of #10818 (M4e), #10524, epic #10486.

M4b's StrongARM latch (sar_comparator.t27, multi-node Newton) run on the PSP103 current tables of M3b at the golden deck's 50 C, judged against ngspice-47 + PSP103 with the golden deck's own dynamic_comparator and dynamic_comparator_top. No change to M4e's numbers.

Read first: the latch does not have the threshold M4e fits

The issue asks the PSP103 latch to predict M4e's 6.7 mV decision offset. It cannot, because the isolated latch has no threshold: the circuit is mirror-symmetric (pair XM3/XM4, pull-ups XM8/XM9 and XM6/XM7, cross-coupled XM10/XM11, resets XM12/XM13, loads), so its flip point (the VDP - VDN at which dp and dn trade places) is zero. Measured:

flip point (VDP - VDN)
model, tt 50 C tables +0.12 uV (2 units of 2^-24 V, rounding)
ngspice PSP103, golden dynamic_comparator, charge swamped or real within +-0.12 uV of 0 (decides right at +-2 units; at d = 0 dp and dn agree to 1e-6 V)
ngspice PSP103, real charge, dynamic_comparator_top (inverter on dp and dn, NAND "Valid") -0.153 mV (dn wins at d = -2570 units, dp at -2540; tight tolerances, reltol 1e-8)
M4e fitted offset +6.7 mV
M4e: latch decision point read from the golden run +2.5 .. +3.1 mV

The top's -0.153 mV comes from the NAND's two inputs not being the same load (the inverters' gate-drain capacitance feeds it back to dp and dn); it favours dp, the golden run's reading is positive, so even the sign differs. Valid rises 897.7 ps after the clock for +2 mV and 920.6 ps for -2 mV. The converter's offset is therefore not the DC threshold of the latch. What an offset of this size IS, in this circuit, is a pair imbalance (XM4, the device whose gate is VDP, at m = 35 or 34 instead of 36; the bulk-tied deck, charge swamped):

ngspice model model vs ngspice
m = 35 +4.2665 mV (dn wins at d = 71579 units, dp at 71581) +4.176 mV (70064) -2.1 %
m = 34 +8.689 mV (145772 / 145774) +8.513 mV (142816) -2.0 %

So the threshold is predicted within 2.1 % as a function of the imbalance, and 6.7 mV is an imbalance of 4.4 % of the pair's current (m of XM4 = 34.42 of 36): a mismatch of the pair or of the input stages, plus whatever the clock edge does (M4e: kickback of 12 and 25 mV in 2 ns), not the latch.

What is in the PR

  • specs/analog/sar_comparator_psp.t27 (module AnalogSarComparatorPsp, 13 tests, 0 vacuous, sealed): the unit conversion, the runs, the flip-point bisection, the pinned ngspice numbers, the tests.
  • specs/analog/mos_psp103_50c_data.t27 (module AnalogMosPsp10350cData, 3 tests, 0 vacuous, sealed): M3b's two tables at 50 C (data and pins; fold_arr and monotone are used, not copied) plus the first 200 held-out R2 points per device. M3b's file is not edited.
  • specs/analog/sar_comparator.t27: the solver generalized in its home spec, no copy: sc_stamp_dev_np, sc_stamp_np, sc_iter_np, sc_step_np, sc_run_np take the NMOS table and the PMOS table as two arguments; sc_stamp_dev, sc_stamp, sc_iter, sc_step, sc_run are one-line forms that pass the one table twice. The 20 tests and every pinned fold are unchanged (20/20, 0 vacuous). The real twin (sc_*_f) and sc_board_step keep their one table.
  • Decks in specs/analog/decks/: psp103_tab_tt50.cir (tables at 50 C), sar_cmp_psp_r2.cir (golden dynamic_comparator, charge swamped), sar_cmp_psp_r2b.cir (the same with every bulk tied to its source, and the unequal-pair scans), sar_cmp_psp_real.cir (real charge, bare and dynamic_comparator_top). They read IHP_PDK, IHP_OSDI and SAR_GOLDEN_SUBCKTS (the variable of analog M4d: the SAR ADC logic and asynchronous clock -- gate level from the netlist, against ngspice PSP103 edges (Closes #10754) #10845) from the environment; the golden deck is fetched at run time and not committed.
  • Seals: two saved (AnalogMosPsp10350cData, AnalogSarComparatorPsp), two renewed (AnalogSarComparator, and AnalogSarSwitch, whose closure holds sar_comparator: its spec and gen hashes are unchanged, only the closure entry and the date moved). seal --verify: all hashes MATCH for the four. sar_adc (analog M4e: the whole SAR converter -- our own ngspice + PSP103 codes of the golden deck pinned, a behavioural converter compared conversion by conversion (Closes #10818) #10854) is not on master yet; it imports sar_comparator and needs its closure entry renewed by whichever of the two lands second.
  • Hand-written foreign lines: +0 / -0 (git diff --numstat over cli, bootstrap, tools, scripts is empty). python3 -I tools/dupe_scan.py: no new duplicate body, no known group grew.

The unit, the devices, the temperature

  • Unit. A table entry is Id * dt / C in units of 2^-24 V. M3b's tables have dt / C = 200 V/A (0.1 ps on 0.5 fF); the comparator's charge unit is 0.125 fF at 0.1 ps, so its dt / C is 800 V/A (SC_STEP_V_PER_A). A word of the solver is exactly 4 times an M3b word (sp_scale); the largest, 4 x 1271348 = 5085392 (NMOS) and 3874764 (PMOS), is under the 2^23 the u64 interpolation allows.
  • Device sizes. The table devices ARE the comparator's: sg13_lv_nmos W 1.0 um, sg13_lv_pmos W 3.0 um, L 0.45 um, ng 1. The comparator's devices have m = 6 .. 36, and PSP103 multiplies every current by m exactly, so the only scaling is the multiplier SC_DV_M the solver already applies. No narrow-width or edge effect is assumed, because no width differs.
  • Temperature. The golden deck is at .options TEMP = 50; M3b tabulated 27 C. I tabulated 50 C with M3b's recipe (preferred and cheap): the same ngspice and checkout reproduce M3b's TT_N, TT_P and TT_R2_N word for word at 27 C, so the toolchain is the same. 50 C: saturation current -5.8 % (NMOS), -4.2 % (PMOS), weak inversion x1.6 at Vgs = 0.25 V. Held-out accuracy (first 200 R2 points): 4.97e-7 A (NMOS) and 3.92e-7 A (PMOS), relative at least 100 uA 0.34 % / 0.14 % (M3b: 0.56 % / 0.41 %).

References (all ngspice-47, PSP103, mos_tt, IHP-Open-PDK 21133f21, 50 C)

  • R2 sar_cmp_psp_r2.cir: the golden dynamic_comparator as it is, with M4b's lumped node capacitances times 1e6 and the step times 1e6 (the same dt / C; PSP103's own charge, a few hundred fF, is 1e-4 of the lumped one), fixed-step backward Euler on the model's grid. Compares the CURRENTS alone. Scales 1e4 / 1e5 / 1e6: the decision time moved 0.83 % then 0.09 %.
  • R2b sar_cmp_psp_r2b.cir: R2 with every bulk tied to its source, the one thing a 2-D table Id(Vgs, Vds) models. The golden deck ties NMOS bulks to ground and PMOS bulks to VDD, so the input pair (source net1, 0.15 .. 0.4 V above ground) and XM8/XM9 carry a body effect the tables cannot have.
  • R1 sar_cmp_psp_real.cir: all of PSP103's charge, an ideal 1.6 ps clock, ngspice's own steps; Xb the bare latch (4 fF on dp, dn), Xtop dynamic_comparator_top.

Sign and times (steps of 0.1 ps after the clock's 50 % point; model / R2b / R2)

decision: model R2b error R2 (golden) error resolution: model R2b error R2 error sign
+-2 mV 53.502 53.854 -0.65 % 57.198 -6.46 % 71.856 72.309 -0.63 % 74.476 -3.52 % right
+-0.5 mV 54.621 54.937 -0.58 % 58.442 -6.54 % 81.255 81.726 -0.58 % 83.462 -2.64 % right

Also 0.125 mV, 31 uV, 10 uV, 1 uV: decision 55.00, 55.13, 55.16, 55.17 (R2b 55.34, 55.46, 55.48, 55.49); at 0.12 uV the sign is right and the resolution is not reached in 140 steps (in ngspice neither). All 14 runs (7 inputs x 2 orientations) decide as R2 and R2b do. Named bounds (from these numbers, both sides): vs R2b decision within 0.66 %, resolution within 0.72 %; vs R2 within 6.7 % and 3.6 %. Nodes at the 8 recorded steps: within 15.9 mV of R2b (the table's error; M4b's square law had 15 mV), 150 mV of R2 (net2 after the decision), 113 mV at dp.

The errors, kept apart

  • table: -0.65 % / -0.63 % (R2b), 15.9 mV.
  • body effect (R2b vs R2, ngspice alone): the golden deck is 6.2 .. 6.5 % slower to decide, 3.0 .. 0.3 % slower to resolve. This is the largest error against the golden deck; the tables carry no Vsb. A 3-D table or a threshold-shift model would take it out. They compose: (1 - 0.0065) / (1 + 0.0621) - 1 = -6.46 % (a test).
  • charge (R1 vs R2): the bare golden comparator with its own charge decides 523.0 ps after the clock at +2 mV, the model 5.35 ps: 96 .. 98 times too fast in the decision, 155 .. 183 times in the resolution. Without charge tables (analog M3c: PSP103 terminal-charge tables, and a weak-inversion-accurate table read, against ngspice #10777) the model's picoseconds are the currents'; the decision TIME is therefore judged against the charge-swamped decks (as the issue's brief anticipated) and only the sign and the flip point against the real one.
  • arithmetic: NOT re-measured: the real twin has no two-table form. Checked instead: last residual at most 3 units (0.18 uV) at all 140 steps of all 14 runs (SP_RES_BOUND; the square law had 2), K = 4 gives the same times (1e-3 of a step in the decision), K = 2 leaves 155 units, K = 1 leaves 20750.
  • discretization: shared with ngspice by construction in R2 and R2b.

Negative controls (tests that pass by seeing the failure)

The decision time is what tells the right tables from the wrong ones; the sign and the flip point cannot, because the symmetric latch decides the right way and flips at zero for any pair of tables (the test asserts that for ss).

  • ss at 27 C (M3b's slow corner) in the tt 50 C run: decision 14.1 % late, resolution 11.5 % late against R2b (bounds 0.66 % / 0.72 %); 7.5 % late against R2.
  • tt at 27 C (right corner, wrong temperature): 5.0 % early / 6.0 % early, 7.5 and 8 times the bounds.
  • PMOS table in the NMOS slot and the reverse: 20.7 % late.
  • Inputs exchanged: the sign check fails. K = 1, K = 2: residual bound broken by 6900 and 52 times.
  • A pin sees one table word changed by one unit.

What it does not claim

Nothing about charge (98 times too fast; the real comparator's time is not predicted). Nothing beyond the measured size of the body effect (6.2 % of the decision time). It does not predict M4e's 6.7 mV (above). The converter's offset (kickback, input stages, delay-cell clock, DAC settling) is not here, and the flip point of the isolated latch says nothing about it. The top's -0.15 mV is measured, not modelled. Other corner or temperature than tt 50 C, other W, L or supply, noise, Monte Carlo mismatch (the imbalance is a deterministic m). The arithmetic error against a real twin. A board, power, or speed against SPICE. The rest value of net1 (0.401 V) is an initial condition shared by the decks and the model, not PSP103's own (an operating point puts floating net1 at 0.71 V on gmin). The ngspice numbers are one run's (ngspice-47 Homebrew, SPARSE 1.3, run on this Mac, 2026-10-19).

Checks

  • t27c test-report: sar_comparator_psp 13 of 13, vacuous passes 0 of 13; mos_psp103_50c_data 3 of 3, 0 of 3; sar_comparator 20 of 20, 0 of 20 (unchanged tests).
  • t27c seal --save then --verify on the four specs: all hashes MATCH. t27c check: 0 errors.
  • Not done, by instruction: gen/, docs/reports/t27b_expectations.json (the two new specs are unlisted; the steward lists them), .claude/, mos_table.t27, mos_psp103*.t27 (no change was needed there; hold_stats is used from mos_psp103, which couples this file to its signature).

Heads-up for #10845 (M4d)

The SAR_GOLDEN_SUBCKTS line in its sar_logic.t27 header, grep -v -i -E '^.(end|global)', also drops the .ends lines (. is any character, so .ends matches .end). The extracted file then has 16 .subckt and 0 .ends, and ngspice stops with "Mismatch of .subckt ... .ends". The line I use in the decks is grep -v -i -E '^\.(end|global)([[:space:]]|$)'.

Reproduce

curl -sfL -o top.spice https://gitlab.symbioticeda.com/Juan_25/uberwave/-/raw/main/wp2-parser/golden/top.spice; sed -n '/^.subckt tracking_switch/,$p' top.spice | grep -v -i -E '^\.(end|global)([[:space:]]|$)' > $SAR_GOLDEN_SUBCKTS; then IHP_PDK=... IHP_OSDI=... SAR_GOLDEN_SUBCKTS=... ngspice -b specs/analog/decks/sar_cmp_psp_r2.cir (and _r2b, _real, psp103_tab_tt50).

🤖 Generated with Claude Code

claude and others added 2 commits October 19, 2026 12:34
@claude
analog M4b-psp: the StrongARM comparator on PSP103 tables at 50 C aga…
93078dd 
…inst ngspice -- sign, times, flip point, imbalance sensitivity (Closes #10856)

The golden comparator, M4b's multi-node Newton solver reused with the NMOS and PMOS tables as two
arguments (sc_*_np; the old names delegate, every old pin unchanged), on the tt 50 C PSP103 tables
(mos_psp103_50c_data.t27, deck psp103_tab_tt50.cir). Judged against ngspice-47 + PSP103 with the
golden dynamic_comparator (charge swamped by 1e6, bulk tied or as in the deck) and with the real
charge, bare and as dynamic_comparator_top. 13 tests, 0 vacuous. The isolated latch has no
threshold of its own: flip point 0 in the model and in ngspice; the top's NAND gives -0.153 mV;
an unequal pair gives the offset, predicted within 2.1 %.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
@claude
Merge master (M4e, #10854) into the PSP103 comparator; reseal sar_swi…
64d4032 
…tch (Closes #10856)

M4e merged first; this branch changes sar_comparator.t27, so sar_switch's closure moved and its seal is
made again on the merged tree. sar_adc does not import sar_comparator (its closure is linear + sar_cdac,
unchanged). sar_comparator 20/20, sar_comparator_psp 13/13, mos_psp103_50c_data 3/3, sar_switch 14/14,
sar_adc 10/10, sar_cdac 13/13, 0 vacuous each; seal --verify all MATCH; no stale closure in any analog seal.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@gHashTag
gHashTag enabled auto-merge (squash) October 19, 2026 05:38
This was referenced Oct 19, 2026
Open
Open
This was referenced Oct 19, 2026
Merged
Merged
Merged
@github-actions

github-actions Bot commented Oct 19, 2026

Copy link
Copy Markdown
Contributor

PR Dashboard

Generated at: 2026-10-19 05:59:43 UTC

Summary

Status Count
Total Open PRs 50
PRs with Failing Checks 36
PRs with All Checks Green 14
READY 2
FAILING 36
PENDING 0
NO CHECKS YET 0

These columns do not partition: 2 + 36 + 0 + 0 = 38, and there are 50 open PRs. A PR is being counted twice or not at all.

Seal Status

  • ⚠️ STALE -- sha256(compiler.rs)=88b8cb73ceda != manifest seal=87e5cbd3ad94.
    The committed NMSE numbers were certified against an older compiler.rs.
    Run scripts/reseal-check.sh locally for the two-step reseal command (advisory; not a merge gate).

This was referenced Oct 19, 2026
Merged
Merged
@gHashTag
gHashTag merged commit 1743da8 into master Oct 19, 2026
31 of 36 checks passed
@gHashTag gHashTag mentioned this pull request Oct 19, 2026
Merged
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

analog M4b-psp: the StrongARM comparator on PSP103 tables -- predict the decision threshold M4e now fits

2 participants

@gHashTag @claude