CryoFlux Loop Architecture Study

Thought for 1 min 50 sec

Rev C is revised in place with the trap-ID and mass reconciliation, fluid-specific failure tables, blackout budgets, and matching nitrogen/argon pressure cases. I’m now checking those deliverables and the archival preservation; the modeled tables do not establish physical trap elimination, safe relief capacity, or a qualified common coupling.

Thought for 3 mins 37 sec

Thought for 24 sec

Awaiting 1 task

"Rev C Documentation Review" completed.

Thought for 31 sec

flux

Verification found one connection-model error: detached pockets in the recovery circuit were routed to the fixed receiver instead of retaining their own recipient. I’m correcting those graph rows and refreshing the linked calculations and audit records; the mass reconciliation and budget calculations passed their checks.

Thought for 12 mins 54 sec

Thought for 17 sec

Thought for 7 sec

Awaiting 1 task

"Loop Architecture Study Planning" completed.

Thought for 55 sec

flux

Revised Rev C in place in Loop Architecture Study in project files, with the four task tables, complete modeled ID/mass reconciliation, blackout budgets, nitrogen/argon cases, and linked calculations and audits; verification found and corrected one detached-recovery routing error.

Corrected Rev B/A were left untouched, with the archived-tail readback limitation disclosed; R-CL10 remains unchanged, D-interface scope remains a Founder question, and no design or release changes were made.

Physical safety and common-interface qualification remain blocked—next priorities are actual valve/cavity drawings and inventories, the approved receiver/pressure basis, and measured parked heat leaks.

Files

Schematic

Layout

3D View

Card-Cage Layout Study

Current study revision C — 2026-10-17. Governing document remains ICD Rev A, 2026-09-20.

Rev C Addendum 2 — chip-face routes under R-CL8A / R-CL8B / R-CL8C

2026-10-17; documentation and budget study only. The complete Founder attachment was read. All prior Rev C, R-CL8A, Rev B and Rev A content is preserved below. No schematic, PCB, component, net, layout or ICD change is made. Article10 remains unchanged. This addendum supersedes only the prior open §8.6 interpretation and the statement that a tile alternative still requires Founder acceptance; it does not retrospectively qualify a material or erase their history.

Founder rulings — verbatim

R-CL8B: “ICD Rev A section 8.6 governs conformal coating of electronic assemblies. The zoned ceramic layers required by R-CL8A are part of the cold sandwich plate construction and are outside the scope of section 8.6. The section 8.6 masking requirement continues to bar conformal coat from connectors and from every thermal interface, including the plate's chip-contact faces. The compliant TIM is retained.”

R-CL8C: “A dense aluminum nitride tile bonded to the chip pedestal by direct-bonded aluminum or an equivalent aluminum-based bond satisfies the R-CL8A chip-contact zone as an alternative to a deposited coating. Tile thickness, bond layers, edge treatment, recess or datum compensation, and thermal penalty are specified in engineering and verified by cold-cycle, dielectric and thermal test before release.”

Closure: R-CL8B closes the historical §8.6 interpretation flag; no ICD revision or further Founder question is required for that closed scope. R-CL8C accepts the bonded-tile construction as an alternative study route, not a selected part or qualified assembly. TIM and all electronic conformal-coat masking remain required.

Evidence and value-classification register

Legend applying to EVERY numerical value in this addendum: I = ICD source value; S = STUDY ASSUMPTION, including user-supplied planning values and inherited unqualified geometry; L = LITERATURE; V = SUPPLIER-STATED; D = DERIVED arithmetic on the declared inputs. Table input and output headers explicitly carry these classifications; case numbers are identifiers. Verified means the cited evidence supports that limited source statement, NOT that our assembly is qualified. N/D = not established, not zero. All planning properties are constant-k screening values; none is cold k(T) qualification.

Table

Input/evidence, class	Value and verification status	Application boundary
I, power and module voltage	CPU358 W (§4.1), OAM750 W /up to900 W (§4.2); module44–59.5 V window inherited from ICD review	Power ceilings, not die heat maps; voltage window is not actual isolation stress or transient specification.
S, inherited geometry	T89 /p98 mm, bank340 mm, hC35 mm, hO35 mm baseline, four/rung hO22 mm equality; local structure/seal4 mm and two6 mm plate bounds	Unqualified geometry, not measured free reserve or pressure-wall permission.
S, planning areas and sensitivities	CPU25 cm², OAM64 cm²; die-scale50 /100 W/cm²; total added datum reserve0.20 mm	Area assumptions, not ICD flux or real die/HBM footprints.
L, Kim2017 measured composite	AD30 µm onAl:122 W/(m·K); GSV30 µm:134; GSV70 µm:117. These are coated-plate, not film conductivities.	5052Al substrate1.5 mm thick and measured bare-Al k138 W/(m·K). Verified source facts, no our-plate qualification.
S, user-supplied deposited-AlN planning base	k20–30 W/(m·K),50 upside; selected finished film caps10 /25 µm	Source does NOT directly state film k20–30. D inverse-series estimates from the preceding L operands are AD30:17.95, GSV30:54.71, GSV70:27.46 W/(m·K), ignoring thermal boundaries and measurement uncertainty. Thus20–30 is a planning band, not a demonstrated AD lower bound. AD and GSV are different processes; no transfer of GSV performance to AD is implied.
V, NTST thermal-spray AlN	Approximately80 W/(m·K),300 V/mil,as-sprayed205–395 microinch roughness; supplier explicitly says characterization is limited.	D:11.811 kV/mm and5.207–10.033 µm roughness conversion. Property test temperature and associated specimen thickness are not specified. Illustrated8 mil coating is D0.2032 mm, not evidence that finished10 /25 µm thermal-spray layers are feasible. Ground/lapped final thickness and minimum dielectric cover TBD.
V, Heraeus AD alumina reference, USER-SUPPLIED attribution	kabout10 W/(m·K), dielectric>18 kV/mm; UNVERIFIED numerical supplier claim in this bounded search: no usable matching datasheet obtained	Exterior reference only, not chip AlN strength or a chip-zone substitution.
V, CeramTec bulk tile reference	AlN170C/170D k170 W/(m·K),20°C,Xe-flash,16×16 mm² specimens,thickness≤3.5 mm;170C dielectric≥15 kV/mm AC/DC at≤1 mm. All values carry±10% measurement tolerance.	Dense bulk reference only;170D dielectric cell blank, not a universal grade guarantee. No coating credit or bonded-assembly guarantee.
S, user-supplied tile/bond planning	AlN0.32 /0.25 mm; Al layers0.3 /0.2 mm; nominal pure-Al k230 W/(m·K); interfaceR''0.5 /2 /5×10⁻⁶ m²K/W per interface	230 is a bulk reference assumption, not the conductivity of a real AlSi wetting/braze/oxide/bond. Actual bond composition, thickness,tolerance,k(T),contact impedance remain unknown.
L, Patterson2012 provenance and inconsistency	Experimental text specifies DBA Al0.3 /AlN0.32 /Al0.3 mm. Table1 instead lists DBA0.3 /0.65 /0.3 mm; it swaps the apparent DBA/AMB ceramic thickness assignments.	Verified source inconsistency, NOT silently resolved. The requested0.92 mm stack is retained as S planning geometry based on experimental prose, not as an unambiguously verified tested thickness.
L, Patterson2012 cycling	Air-to-air shocks−40↔200°C. DBA no observed complete-delamination failure through6000 cycles; reportedMTTF>6000,DBC AlN mean660 cycles.	Tested constructions/patterns differ; not universal material lifetimes. DBA trace curl around3900 cycles was not counted as failure. AMB camber was tentatively linked to patterned-top/full-bottom asymmetry; baseplate attachment could change stress. Neither symmetric nor single-sided tile on ourAl/Cu insert is qualified.
S, thinner tile availability	Al0.2 /AlN0.25 /Al0.2 mm and AlN0.25 +Al0.2 mm	Exact commercial standard combinations NOT ESTABLISHED in bounded lookup; hypothetical coupon geometries, not catalog selections.

The Kim inverse estimate uses D: k_f=t_f/[(t_Al+t_f)/k_composite−t_Al/k_Al]. It cannot separate film and boundary resistances; no confidence interval or production process guarantee follows. It also prevents misreading the composite122–134 values as exceptionally conductive films. The NTST80 row is a separate thermal-spray sensitivity, never evidence of AD80. All actual Al/Cu-substrate cold-cycle,dry/wet dielectric and measured through-plane k(T) qualifications remain open.

A. Finished-datum stack budget

S: cases1/2 use the same10 /25 µm cap per chip face AND per exterior face. Tile cases use 25 µm exterior ceramic per plate, explicitly S, not specified by the Founder attachment. Each selected layer dimension below is a maximum finished budget cap, including its allocated bond/wetting/finish tolerances; actual tolerance feasibility is unverified. A separately added pedestal-attachment layer,bond fillet,edge overspray or lap allowance outside these caps must enlarge this budget. Retain S0.20 mm total datum reserve; it is not physical coating thickness and not proof that any tolerance fits.

For chip stackc,exterior cape and recessd per plate, D: physical external buildup b=2(c−d)+2e; reserved envelope addition a=b+0.20 mm;4 mm consumed=a; residual=4−a. Physical uncompensatedT/p=89+b /98+b; conservative reservedT/p=89+a /98+a. Three/rung uses N=floor[340/(98+a)] and residual340−3(98+a). At hC35, four/rung hO ceiling=22−a; if datum reserve is excluded for physical-only comparison,the ceiling is22−b. These are scalar bounds, not measured packing with tolerances.

Table

Case /S construction and mounting	S chip stackc /exteriore; recessd, mm per plate	D physicalb /reserved additiona, mm	D4 mm consumed /residual, mm	D physical uncompensatedT /p, mm	D reservedT /p, mm	D N; three-pitch bank residual mm	D four/rung hO ceiling incl.reserve, mm
1 deposited10 µm	0.010 /0.010;0	0.040 /0.240	0.240 /3.760	89.040 /98.040	89.240 /98.240	3;45.280	21.760
2 deposited25 µm	0.025 /0.025;0	0.100 /0.300	0.300 /3.700	89.100 /98.100	89.300 /98.300	3;45.100	21.700
3 symmetric0.3Al/0.32AlN/0.3Al,surface	0.920 /0.025;0	1.890 /2.090	2.090 /1.910	90.890 /99.890	91.090 /100.090	3;39.730	19.910
3 symmetric,fully recessed	0.920 /0.025;0.920	0.050 /0.250	0.250 /3.750	89.050 /98.050	89.250 /98.250	3;45.250	21.750
4 single0.32AlN+0.3Al,surface	0.620 /0.025;0	1.290 /1.490	1.490 /2.510	90.290 /99.290	90.490 /99.490	3;41.530	20.510
4 single,fully recessed	0.620 /0.025;0.620	0.050 /0.250	0.250 /3.750	89.050 /98.050	89.250 /98.250	3;45.250	21.750
5 hypothetical thin symmetric0.2Al/0.25AlN/0.2Al,surface	0.650 /0.025;0	1.350 /1.550	1.550 /2.450	90.350 /99.350	90.550 /99.550	3;41.350	20.450
5 thin symmetric,fully recessed	0.650 /0.025;0.650	0.050 /0.250	0.250 /3.750	89.050 /98.050	89.250 /98.250	3;45.250	21.750
5 hypothetical thin single0.25AlN+0.2Al,surface	0.450 /0.025;0	0.950 /1.150	1.150 /2.850	89.950 /98.950	90.150 /99.150	3;42.550	20.850
5 thin single,fully recessed	0.450 /0.025;0.450	0.050 /0.250	0.250 /3.750	89.050 /98.050	89.250 /98.250	3;45.250	21.750

All D additions fit arithmetically inside the nominal S4 mm, including the largest symmetric surface stack. This does NOT establish available reserve. Fixed89/98 is conditional on a dimensioned redistribution proving existing structure/seals≤the row residual while preserving module allocations,TIM,hard stops,guides and sleeve. Without that redistribution or finished-datum machining, none remains exactly89/98; the table reports its growth. No case is demonstrated to fit only with recess, and no case fails the nominal4 mm arithmetic. Actual pass/fail is blocked by absent local structure/seal drawings. Fully recessing the chip tile still leaves exterior0.050 mm physical plus0.20 mm reserved addition across the sandwich; recess alone does not restore89/98. With proven compensation,base three/rung remains unchanged; restoring the hO22 equality additionally requires complete compensation of all applicable additions,not just chip ceramic.

Recess material/removal ledger (all dimensions S, feasibility UNVERIFIED): case3 requires0.920 mm,case4 0.620 mm,thin symmetric0.650 mm,thin single0.450 mm removal from each standalone pedestal under the tile. On anAl pedestal the removed material isAl; on the inherited solidCu pickup/insert it isCu; if the footprint crosses both,an engineeredAl/Cu step/joint must be defined. Removed volume=depth×actual machined footprint,not known here. Remaining pedestal stock,insert embedment and distance to channels/seals are N/D. Pressure-wall material removed is ZERO by constraint: no shaving of the S6 mm pressure-wall bound is permitted. If no independent stock exists,full recess is blocked,not assumed machinable. At every affectedAl/Cu insert joint,pedestal ligament,seal/weld neighborhood,clamp boss and channel ligament,recheck pressure boundary,bulge,fatigue,stiffness,CTE,bond integrity and finished plane; pressure reanalysis is mandatory wherever those features participate in the boundary. No machining credit without a released section drawing and analysis/test. Internal wetted lining still consumes channel/port/seal dimensions,not external width except at protruding flanges/transitions; wet-zone budgets below remain open.

B. Gross added chip-path thermal resistance and temperature rise

I:358 /750 /900 W as registered above. S:25 /64 /64 cm² pickup areas. D fluxes respectively 14.32 /11.71875 /14.0625 W/cm² =143200 /117187.5 /140625 W/m²; none is ICD-specified flux. S die-scale sensitivities50 /100 W/cm² correspond to D500000 /1000000 W/m². Screening formula D R''=Σ(t_i/k_i)+nR''_interface, ΔT=q''R''. Exterior ceramic is normally outside this chip-to-coolant series path. CPU crosses only its own chip layer; OAM only its own opposite layer,never both in series.

All R'' numbers in the following tables are D in units of10⁻⁶ m²K/W; all ΔT values are D kelvin. Film thickness,k and tile interface sensitivities are S planning inputs with evidence boundaries from the register. No TIM,spreading,wet-coating,convection or extra pedestal-attachment resistance is included in these gross chip-layer results.

Table

Case /S film thickness µm	S k W/(m·K),process interpretation	D R'' ×10⁶	D ΔT CPU358/25	D ΔT OAM750/64	D ΔT OAM900/64	D ΔT50 W/cm²	D ΔT100 W/cm²
1 /10	20 AD planning base	0.500000	0.07160	0.05859	0.07031	0.25000	0.50000
1 /10	30 AD planning base	0.333333	0.04773	0.03906	0.04688	0.16667	0.33333
1 /10	50 upside only	0.200000	0.02864	0.02344	0.02813	0.10000	0.20000
1 /10	80 separate thermal-spray sensitivity,thin-process feasibility UNVERIFIED	0.125000	0.01790	0.01465	0.01758	0.06250	0.12500
2 /25	20 AD planning base	1.250000	0.17900	0.14648	0.17578	0.62500	1.25000
2 /25	30 AD planning base	0.833333	0.11933	0.09766	0.11719	0.41667	0.83333
2 /25	50 upside only	0.500000	0.07160	0.05859	0.07031	0.25000	0.50000
2 /25	80 separate thermal-spray sensitivity,thin-process feasibility UNVERIFIED	0.312500	0.04475	0.03662	0.04395	0.15625	0.31250

For tiles, S bulk-referencek_AlN170 and nominal bulk-Alk230 W/(m·K). Ideal series compositek=(total layer thickness)/(Σt_i/k_i), excluding all interfaces, is D204.852 forcase3,194.559 forcase4,202.510 for thin symmetric,192.295 for thin single W/(m·K). These are effective ideal composites withAl layers,not ceramic k and not actual assembly measurements. Corresponding layer-onlyR'' are D4.491049 /3.186701 /3.209719 /2.340153×10⁻⁶ m²K/W.

Define n as the number of internal Al/AlN thermal bond interfaces: n2 symmetric, n1 single-sided. The interface sensitivity is per such interface,not per tile. A symmetric prefabricated tile's bottomAl-to-pedestal attachment is an additional OPEN resistance, absent from the two-interface count; add its layer thickness/contact resistance explicitly. For single-sided construction,the sole Al/AlN interface is counted,but Al-layer-to-pedestal attachment remains additional/open unless a demonstrated integral bond model explicitly includes it. Do not claim either result is the full bonded-to-pedestal impedance. Real alloy/wetting-layer k may also invalidate the pure-Al reference.

Table

Case (S stack),surface AND recess use same gross thermal row	S n;per-interfaceR'' ×10⁶	D totalR'' ×10⁶	D ΔT CPU	D ΔT OAM750	D ΔT OAM900	D ΔT50 W/cm²	D ΔT100 W/cm²
3 symmetric0.92 mm	2;0.5	5.491049	0.78632	0.64348	0.77218	2.74552	5.49105
3 symmetric	2;2	8.491049	1.21592	0.99504	1.19405	4.24552	8.49105
3 symmetric	2;5	14.491049	2.07512	1.69817	2.03780	7.24552	14.49105
4 single0.62 mm	1;0.5	3.686701	0.52794	0.43204	0.51844	1.84335	3.68670
4 single	1;2	5.186701	0.74274	0.60782	0.72938	2.59335	5.18670
4 single	1;5	8.186701	1.17234	0.95938	1.15125	4.09335	8.18670
5 hypothetical symmetric0.65 mm	2;0.5	4.209719	0.60283	0.49333	0.59199	2.10486	4.20972
5 thin symmetric	2;2	7.209719	1.03243	0.84489	1.01387	3.60486	7.20972
5 thin symmetric	2;5	13.209719	1.89163	1.54801	1.85762	6.60486	13.20972
5 hypothetical single0.45 mm	1;0.5	2.840153	0.40671	0.33283	0.39940	1.42008	2.84015
5 thin single	1;2	4.340153	0.62151	0.50861	0.61033	2.17008	4.34015
5 thin single	1;5	7.340153	1.05111	0.86017	1.03221	3.67008	7.34015

Recess does not remove ceramic thermal resistance. Above rows show gross added layers; net change versus the old metal pedestal would additionally subtract R''_removedmetal=d/k_actualmetal only for an actually removed equivalent heat path. No such credit is calculated because alloy/Cu grade,k(T),actual footprint,spreading and available standalone stock are unestablished. Added attachment/TIM/contact,metal spreading/insert joint,internal wetted barrier and coolant-film terms remain separately OPEN. Identical footprint/one-dimensional conduction is not a validated model of die/HBM hotspots or lateral metal layers. No chip Tcase/junction release follows from these ΔT values.

C. Dielectric planning ratios — not allowable working voltage or hipot

D ratio=(reference strength×ceramic thickness)/comparison voltage. I comparison window44 /59.5 V only; it is NOT evidence that either is the maximum isolation stress. All hipot voltage,waveform,duration,IR/leakage criteria,transients,clearance/creepage and safety review remain TBD. Bulk-field×thickness is used here solely for requested planning ratios,not asserted assembly breakdown voltage. No insulation credit is given toAl layers,TIM or nominal untested finish.

Table

Case and process /status	V or S strength input and S ceramic thickness	D ratio vs44 V /59.5 V	Restriction
1 AD AlN10 µm	Actual strengthN/D	N/D /N/D	Do not inherit NTST spray or bulk strength.
2 AD AlN25 µm	Actual strengthN/D	N/D /N/D	No dielectric approval from thermal k.
1-like thermal-spray sensitivity10 µm,NOT demonstrated process	S rounded11.8 kV/mm from NTST V claim;0.010 mm	2.682 /1.983	Conditional calculation only at unverified finished thickness; not AD.
2-like thermal-spray sensitivity25 µm,NOT demonstrated process	Same S11.8;0.025 mm	6.705 /4.958	Finishing may expose pores or remove minimum insulation.
3 symmetric and4 single,tile,surface/recess	V bulk-reference15 kV/mm;S AlN0.320 mm	109.091 /80.672	Bulk planning reference,not bonded/edge assembly guarantee; Al thickness adds zero dielectric margin.
5 thin symmetric/single,surface/recess,hypothetical	V bulk-reference15;S AlN0.250 mm	85.227 /63.025	Same restriction; exact thin commercial stack unverified.
Tile exterior cap,AD alumina reference ONLY	USER-SUPPLIED V attribution>18 kV/mm UNVERIFIED;S0.025 mm	Conditional >10.227 />7.563	Not chip strength; reference only,no assembly approval.

The symmetric chip-side outerAl layer contacts TIM and is electrically floating/isolated from the plate unless an expressly engineered isolation topology states otherwise. Keep no metallic fillet,edge wrap,fastener,grinding smear,swarf orCu/Al seam bridging aroundAlN; inspect minimum ceramic thickness,edge setback/creepage,defects,pinholes,chips and isolation after clamp/thermal cycling. A floating plate still needs charge/field and accessible-conductor review. Single-sidedAlN also needs edge protection and minimum thickness after finishing. R-CL8C's alternative does not waive these checks.

D. Planning recommendation and honest fit classification

No production part/process selected. Start parallel coupon planning with case1/2 depositedAlN and case3 symmetric tile as the literature-backed construction lead; case4 tests lower gross thermal/geometric burden but carries extra asymmetry/camber/cracking uncertainty. Include hypotheticalcase5 only if a supplier can provide exact drawings/process/certification; do not describe it as standard. Patterson's long-shock evidence supports investigatingAl-based bonds,not copying a lifetime into ourAl/Cu pressure plate; the conflicting published thicknesses must be resolved before claiming the exact0.92 mm stack has that test pedigree. Its patterned metallization,plating,roughening and failure definition are not our flatTIM face.

Fit categories: proven89/98 without compensation: NONE. Conditional within original4 mm through explicit local structure/seal redistribution: ALL listed cases, including surface symmetric withD1.910 mm remaining. Demonstrated “only fits with recess”: NONE established. Demonstrated nominal4 mm nonfit: NONE. Fully recessed options consumeD0.250 mm of the local allowance versus surface tileD1.150–2.090 mm,but require independent pedestal stock and still need exterior/datum accommodation. Recess machining practicality,pressure-wall availability and seal redistribution are not assumed. Three/rung remains numerically possible for every uncompensated reserved envelope; no new module-height proof or four/rung promise is made.

E. Coupon matrix and retained provisional gates

The earlier R-CL8A qualification plan remains in force as S provisional engineering targets,not released ICD requirements. Every coupon must represent the actualAl alloy/pedestal andCu insert joint,edge andjoining/finish sequence; include flat witnesses,actual joint coupons and eventually sealed pressure plates. The population rule remains S≥5 coupons per process/substrate/thickness/zone andS3 complete plates,not statistical lifetime qualification.

Table

Coupon family,all proposed S	Surface/recess pairing and datum requirement	Additional gate emphasis; all actual results UNVERIFIED
DepositedAlN10 and25 µm cases1/2	Surface films on actualAl/Cu/pedestal transition; optional precompensated finished-datum counterpart with removal explicitly controlled	Measure through-plane k(T),minimum dielectric thickness,adhesion,residual stress,porosity,seam coverage,roughness and fullTIM/clamp impedance; no bulk property transfer.
NTST-type ground/lapped thermal-sprayAlN	Finished thickness TBD by process supplier,not automatically10 /25 µm; paired surfaced and compensated/recessed coupon only after stock review	Record as-sprayed,removed and minimum finished layers; dielectric/pinhole andthermal tests after grinding; verify phase/oxidation,bondcoat and edges. A newly chosen finished thickness needs a new budget/table.
AD alumina reference	Exterior zone on actualAl/Cu/edges,not a replacement for chipAlN	Confirm user-attributedk10 and>18 dielectric values independently; humidity,abrasion,masking,adhesion. Coolant tests only if that exact recipe is proposed for a wetted zone.
Symmetric0.3Al/0.32AlN/0.3Al	Surface vs full0.920 mm standalone-pedestal recess,if available	OuterAl isolation,edge bridges,bottom attachment impedance,flatness/TIM face roughening,pressure/load path andCTE fatigue. Resolve literature thickness discrepancy.
SingleAlN0.32+Al0.3	Surface vs full0.620 mm standalone-pedestal recess,if available	Camber,brittle cracking,clamp bending,Al/AlN andpedestal bond impedances,edge dielectric andminimum thickness.
Hypotheticalthin symmetric0.2/0.25/0.2 andsingle0.25+0.2	Surface vs0.650 /0.450 mm recess respectively,subject to exact supplier feasibility	Do not assume commercial standardness or machinability; measure warpage,handling yield,edge chips,min thickness and bond stack.

Retain all earlier gates,with explicit S provisional counts:100-cycle feasibility,1000 routine cycles−20°C↔actual hot-end,500 qualification cycles−40°C↔actual hot-end; include I+5/+7°C admission/release behavior,not a substitute for service hot-end. S adhesion retention≥80% with zero through-barrier cracking/delamination/leak; absolute strength remainsTBD. S thermal-impedance drift≤10% plus actual module temperature limits; measure atI−40/−20°C modes,S+7°C test point andactual hot-end at controlledTIM/clamp load. Apply I powers over S pickup maps and S50/100 W/cm² sensitivity loads with actual thermal source/contact maps before release. Dielectric IR/hipot/edge tests remain before/after cycling,wet/dry where actual exposure occurs,levelsTBDreview,not arbitrary certification.

Retain S1000 h circulated coolant exposure withS100/250/500/1000 h checkpoints,on actual wetted ceramic zones andrepresentativeAl/Cu seams,including joined/sealed plate coupons,not artificial immersion of every nominally dry chip face. Dry chip/tile faces get humidity/condensation and service recovery exposure; direct coolant exposure there is added only if the actual leak/exposure case requires it. Retain S100 purge-service and100 controlled humidity/condensation excursions withreal facilityRH/dwell/dew-point inputsTBD. Nitrogen/dew-point interlocks are not replaced by exterior ceramic. Pinhole/holiday inspection,pressure/leak/flow/channel-reduction testing,galvanic-current/corrosion monitoring andpost-exposure electrical tests remain mandatory; coolantchemistry/pressure/flow/hot-end andnumeric acceptance limits remain open. Add post-machining pressure-ligament/stiffness checks,edge dielectric inspection,flatness/roughness and bond/attachment cross-sections. All gates must pass on the actual joined,sealed,clamped assembly; supplier/literature claims alone do not qualify it.

Addendum 2 verification and scope receipt

A–E completed as documentation/calculation/evidence-status study. Calculation receipt: native matrix-evaluation attempts failed (assignment unsupported,then range limit); bounded Python calculation via the project-computer calculation path succeeded with exit code0, script /root/work/cage-study/addendum2.py. It evaluated every stack/reserve/physical and reservedT/p/bank count/hO ceiling,all film andtile/interfaceΔT rows,ideal seriesk,dielectric ratios,NTST unit conversions andKim inverse estimates. This is analytical1-D screening,not FEA orhardware qualification. Read-only full-document readback confirmed the entire new Addendum2, every new inline citation marker, both verbatim rulings and the historical closure annotation persisted. The displayed response truncates the long historical archive,not the new section; no new line-by-line review of all archived revisions is claimed. The old archive is preserved by additive patch,not replaced. No electrical mutation means no changed-net ERC/DRC applies; existing electrical issues andArticle10 gate remain untouched.

Evidence gaps /blocked release items: exact actual4 mm structure/seal reserve,module contact andfinished datums,available standaloneAl/Cu pedestal stock,pressure/camber/fatigue/stiffness,bondalloyk(T) andattachmentR'',qualified finishedfilm thickness/k_z(T),ADAlN strength,Heraeus numericaldatasheet,exactthintile commercial combinations,Patterson as-tested thickness correction,coolant/pressure/hot-end/isolationstress andcomplete assemblycold qualification. Unattempted by design: procurement,coupon fabrication,machining,FEA,hardwaretest,partselection,schematic/PCB/ICD mutation. Completion of this assigned batch does not complete the overall architecture or release the design. Earlier dry-mass comparisons remain uncoated references; ceramic/bond/removedmetal masses are not silently updated.

Rev C addendum — Founder ruling R-CL8A: zoned ceramic sandwich plates

2026-10-17; user-approved STUDY REQUIREMENT, documentation only. Preserve all earlier Rev C/B/A content below. R-CL8A requires EACH sealed sandwich plate to have three distinct ceramic zones: (1) internal wetted channels: continuous ceramic barrier isolating BOTH aluminum plate and copper die inserts from coolant for corrosion/galvanic protection; (2) chip-contact faces: thin high-conductivity AlN-class dielectric, NO anodic oxide at these contacts; (3) exterior: condensation-resistant dielectric ceramic. No polymer-only replacement is accepted. This records the Founder ruling but does not independently amend governing ICD Rev A, approve architecture, qualify a coating, or authorize schematic/PCB changes.

The entire ICD Rev A was read again for this addendum. Articles1.1/2.2 sealed dry nitrogen and no-air/conduction doctrine, §§9.1–9.4 metal clamp paths/compliant TIM/hard stops, and Article10 gates remain intact. Warning—interpretation conflict to close: §8.6 says “Conformal coat: low-modulus silicone class with verified thermal shock; connectors and thermal interfaces masked.” R-CL8A is recorded here as plate-zone ceramic processing, not a silent replacement of that electronics coating rule; clarify whether chip-side ceramic beneath the replaceable TIM is outside §8.6's masking scope. If not, Founder-approved ICD revision is required. This study neither removes the compliant TIM nor coats connectors. The exterior dielectric does not establish that condensation cannot occur and does not replace dry purge, dew-point sensing or door-open recovery.

Historical flag closure annotation — 2026-10-17: The §8.6 interpretation warning in the preceding paragraph is now CLOSED by R-CL8B, quoted verbatim in Rev C Addendum2 above; its conditional ICD-revision escalation no longer applies to this scope. The original paragraph remains as history. R-CL8C also supplies the previously requested Founder acceptance of the bonded-tile alternative,subject to engineering andqualification; the old selection conclusion below remains historical,not a new approval request.

A. Finished-datum stack and bounded 4 mm redistribution

Case6 inherited unqualified baseline is 3 mm carrier +35 mm COMPLETE CPU face allocation +35 mm COMPLETE OAM face allocation +two6 mm pressure plates +4 mm combined local structure/sealing allowance = 89 mm sled; guide5 mm+sleeve4 mm gives 98 mm pitch. The two35 mm allocations must remain available to the modules, external compressed TIM and pedestal/tolerances as originally defined. The4 mm is NOT proven free reserve: its actual structure/seal commitments have never been released.

From each chip toward its own coolant path: compliant TIM/contact → inward AlN-class chip coating → metal pedestal/insert and pressure wall → internal wetted ceramic → coolant. Opposite the metal on the outward dry surface is exterior ceramic. These are three zones on each plate, not six serial thicknesses across the sandwich. CPU heat crosses its own chip coating/plate/wet coating; OAM heat crosses its own opposite plate. Neither heat source crosses both chip coatings in series. Where a local metal feature lies between an inward and outward finished datum, both coatings count geometrically; wetted films lie inside cavities and normally consume hydraulic dimensions rather than add external thickness.

For a conservative uniform cap t on EACH inward chip coating and EACH exterior coating, the uncompensated external sandwich increment is 2t_chip+2t_ext =4t. Treat t as the MAXIMUM finished local buildup, including any associated adhesion/bond/interlayer and minimum-thickness compliance after finishing; this is not a nominal coating with an uncounted positive tolerance. Nonuniform overspray, edge buildup and seam transitions must stay within the cap. The following is a proposed bounded redistribution, not a released process tolerance: separately reserve another 0.20 mm total for added coating-related datum/flatness/assembly uncertainty; the existing TIM/compression tolerances must already be inside hC/hO, not borrowed again. Any extra bond layer or tolerance outside these caps requires a new budget.

Table

Maximum per chip/exterior zone	Two chip +two exterior buildup	Added datum reserve	Total moved out of existing4 mm	Residual for ALL existing local structure/seals	If no reserve/machining is available: new T /p	Four/rung hO ceiling with hC35, no redistribution
50 µm	0.20 mm	0.20 mm	0.40 mm	3.60 mm	89.40 /98.40 mm	21.60 mm
100 µm	0.40 mm	0.20 mm	0.60 mm	3.40 mm	89.60 /98.60 mm	21.40 mm
150 µm	0.60 mm	0.20 mm	0.80 mm	3.20 mm	89.80 /98.80 mm	21.20 mm

Conditional answer only: 89/98 mm can remain unchanged if a dimensioned design demonstrates existing local structure+seal needs are no larger than3.60/3.40/3.20 mm respectively AND frees0.40/0.60/0.80 mm within that original4 mm, or if qualified pre-machined recesses/finished datums compensate the coating buildup while retaining pressure strength and all tolerances. There is presently no drawing proving either route. Do not shorten the35 mm faces, reduce hard-stop/TIM compression, seal lands, guides or independent sleeve allowance to force fit. Changing pedestal metal height under its coating must maintain the finished chip-contact plane and original module/TIM clearance. A6 mm plate is an unqualified bounding slab, NOT permission to shave a pressure wall; local wall/insert/seam strength, pressure bulge, fatigue, alloy temper, weld and clamp paths need re-analysis before machining credits.

Without compensation three/rung still fits nominally (3×98.8=296.4 mm within340 mm at the largest trial), but the equality hO22,p85,four/rung has ZERO bank reserve and cannot absorb even50 µm coatings silently. Without the0.20 mm datum reserve, pure film-only hO ceilings would be21.8/21.6/21.4 mm; use the more conservative table if that reserve is retained. Finished-datum compensation must include inward coatings as well as exterior films, not just subtract paint from an outside dimension.

Wetted thickness tw on opposed channel walls gives finished h=h0−2tw and w=w0−2tw for a rectangular local section. At tw50/100/150 µm each dimension loses 0.10/0.20/0.30 mm; illustrative1×1 mm channels would lose19/36/51% area, respectively (not the actual channel design). Recompute hydraulic diameter, roughness, pressure drop, flow distribution and heat transfer; do not assume area loss scales pressure drop linearly. Pre-machining channel dimensions larger by2tw can preserve finished flow area only if pressure-wall/ligament/insert-joint requirements still pass. PEO converts some parent metal with process-dependent inward/outward growth; use measured finished dimensions, not an assumed universal growth fraction. Wet-zone coatings are not added to the four-film external sum unless they physically project at a seam, port, seal or datum.

B. Thermal penalty: exact ICD basis versus assumed area

No CPU/OAM heat flux or chip-contact area is explicitly stated anywhere in ICD Rev A. Exact power wording: §4.1 “358 W module input ceiling”; §4.2 “AMD Instinct-class (750 W, 54 V) and Intel Gaudi-class (up to 900 W) envelopes until module selection.” Article3.3's1.5/1.8 kW is the entire slot input envelope, NOT CPU/OAM die flux. The earlier study's300 W CPU/900 W nominal OAM/1000 W peak OAM allocations were study assumptions;1000 W is not §4.2's ratified envelope and is not relabeled as an ICD value here.

For calculation ONLY assume uniform effective pickup areas inherited from the hybrid insert concept: CPU50×50 mm=25 cm² and OAM80×80 mm=64 cm². These are NOT actual die/HBM areas or vendor contact maps. Conservatively allocate the complete module input power as heat to its single stated pickup for this first-pass coating calculation; in reality VRMs, HBM and other pickups split heat, and hotspot flux may be higher. Insert area is not evidence of heat spreading to that entire area at the coating plane. Thus the following fluxes are DERIVED FROM ICD POWER WITH ASSUMED AREA, NOT ICD FLUX.

For one uniform chip-face layer, steady1-D screening only, not FEA:

q''=P/A;\qquad R''_{coat}=t/k_z(T)\ [m^2K/W];\qquad \Delta T=q''R''_{coat}

For variable conductivity, use measured through-plane k over the actual local temperature interval; the constant-k values below are declared sensitivities, not an extrapolated k(T) qualification. At ASSUMED k_z=50 W/(m·K), R'' at50/100/150 µm is 1/2/3×10⁻⁶ m²K/W =0.01/0.02/0.03 cm²K/W.

Table

Heat source /ICD article	Power used	Assumed pickup area	Derived flux W/cm² /W/m²	ΔT at50 µm,k50	ΔT at100 µm,k50	ΔT at150 µm,k50
CPU §4.1 input ceiling	358 W	25 cm²	14.320 /143,200	0.1432 K	0.2864 K	0.4296 K
AMD-class OAM §4.2	750 W	64 cm²	11.71875 /117,187.5	0.1172 K	0.2344 K	0.3516 K
Gaudi-class OAM §4.2 upper envelope	900 W	64 cm²	14.0625 /140,625	0.1406 K	0.2813 K	0.4219 K

All table numbers scale by50/k_z: k10 →×5, k100 →×0.5, bulk-reference k170 →×0.2941. At150 µm CPU ΔT is2.148/0.4296/0.2148/0.1264 K for k10/50/100/170;900 W OAM gives2.1094/0.4219/0.2109/0.1241 K. These10/50/100 values are engineering sensitivity assumptions, not a demonstrated minimum or guaranteed film range at−40°C. A poor film with k1 would give21.48 K CPU at150 µm; the sensitivity is not a lower-bound proof. Smaller real pickup area multiplies flux and ΔT proportionally. No Tcase/junction safety margin can be approved without actual maps, plate/coolant temperatures and measured full thermal impedance.

Validated material evidence—not transferable assembly guarantees: CeramTec lists bulk Alunit AlN170C/170D/HP 170 W/(m·K) at20°C, Xe-flash on16×16 mm² specimens, material-specific thickness≤3.5 mm. Its values carry±10% measurement tolerance. This is a dense ceramic substrate reference, not deposited AlN onAl/Cu; the source does not establish our film's through-plane k(T).

Perez et al., author-hosted ACS Nano2023 paper and supplement, reports room-temperature CROSS-PLANE approximately36–104 W/(m·K) in approximately600 nm films onSi(111)/c-Al₂O₃, not on our plate alloys/inserts. Deposition was below100°C. That is real process-specific thin-film evidence, but it is NOT proof of50–150 µm feasibility or of k at−20/−40°C. S11's685 nm,75%N₂ sample is identified, but the retrieved S11 evidence is capacitance characterization, not a numerical breakdown-field or cold-cycle qualification.

For temperature dependence, Xu et al., JAP2019 author PDF, measured crystalline bulk AlN674±56 at100 K,237±6 atroom temperature,186±7 W/(m·K) at400 K. Heat flow is perpendicular to the top sample surface. This establishes that k depends onT and material quality, not a usable coating k at233/253 K; no linear interpolation or beneficial cold correction is credited. Actual deposited film k_z(T),233 K through hot-end, remains unverified.

TIM/contact resistance, coating-metal thermal boundary resistance, any bond/adhesion layer, insert-to-Al seam/spreading resistance, wet-coating resistance and liquid-film convection are separate OPEN series terms on each actual heat path. Do not fold unknown bonds into the quoted AlN k or count them twice. In area-normalized form R''_pickup=R''_TIM/contact+R''_chipcoat+R''_bond/boundary+R''_metal/spreading+R''_wetcoat+R''_liquid, with area transformations/spreading solved when areas differ. Wetted ceramic can dominate if it has low k; its unknown area and thickness prevent a numeric total penalty. The exterior coating is normally not in the primary chip-to-coolant path. This is why the sub-kelvin chip-film calculation alone does not confirm a qualified cold plate.

C. Bounded process survey and zone applicability

Evidence statuses below distinguish validated PDF facts from scraped official supplier/publisher leads. UNVERIFIED for actual assembly means no qualified combination of process,Al alloy/Cu grade,interlayer,thickness,geometry,humidity/coolant and thermal cycles; it does not mean the technology is impossible. No supplier's service-temperature headline or bulk breakdown strength is accepted as cold-cycle evidence.

Table

Candidate process /zone	Evidence and realistic thickness/process limit	Dielectric strength status	Actual Al/Cu sealed-plate cold cycling
Reactive sputtered/PVD AlN,chip zone (not anodic)	Validated Perez cross-plane measurements above. Demonstrated submicrometer films; growing50–150 µm with retained stress,adhesion,flatness and throughput is UNVERIFIED. CVD/PECVD AlN is an alternative development route,not an automatically equivalent process; high-temperature crystalline routes must not be assumed compatible withAl temper,Cu/Al joint or seals.	Actual film breakdown at selected thickness UNVERIFIED; S11 title alone is not proof. Bulk AlN170C/HP≥15 kV/mm AC/DC at≤1 mm is a bulk substrate fact only, not a field or standoff guarantee for deposited ceramic.	UNVERIFIED; no applicable−40↔hot-end cycle count/result established.
Aerosol-deposited pureAlN,chip zone	2012 publisher abstract reports room-temperature denseAlN onAl and>88 W/(m·K),but direction/process details were not PDF-validated here. 2016 paper DOI is a thick-film-onAl lead; abstract repositories suggest up to30 µm,but exact range remains UNVERIFIED. 2024/25 IOP study reports approximately10 µm onSi in accessible text,not ourCu inserts. Publisher PDF download failed; no50–150 µm qualified process is established.	Numerical breakdown for exact pureAlN/Al orCu thickness UNVERIFIED; impact damage,residual stress,edge coverage and post-polish minimum thickness need coupons.	UNVERIFIED. Room-temperature deposition is not thermal-shock/cold-cycle qualification.
PEO/ceramic anodic barrier,wettedAl and/or exteriorAl ONLY; forbidden atchip face	Keronite Chimera official page scraped: approximately50 µm pure ceramic onAl alloy,>3 kV breakdown,>1 W/(m·K). Official-web evidence only,not PDF-validated or our process approval. Its ceramic-composite/sealer variant is separate; do not substitute polymer-only coating or infer ceramic-sealer chemistry. Supplier regime datasheet page is gated; exact100/150 µm regime and growth tolerance UNVERIFIED. PEO onAl does not establish coverage of bareCu inserts.	Supplier's>3 kV atapproximately50 µm is NOT a wet/cold allowable working voltage. Exact alloy,geometry,test electrode,DC/AC and production minimum UNVERIFIED.	UNVERIFIED for sealedAl/Cu assembly. General thermal-cycle claims are insufficient. A separateCu-compatible deposited ceramic and verified continuous seam transition are mandatory.
ALD oxide/nanolaminate ceramic,wettedAl+Cu/exterior,sealing overlayer candidate	Beneq official corrosion page supports conformal barriers andAl₂O₃/TiO₂ stacks as an official-web lead; actual coolant approval UNVERIFIED. Part-coating page scraped lists oxide processes≤300°C andAlN≤380°C,not our approved thermal budget. ALD process explanation describes nanometer-scale growth. ALD is a conformal thin seal/overlayer candidate,not evidence of practical50–150 µm bulk buildup. Nucleation,coverage inside final channels and aqueous durability need validation.	Numerical dielectric breakdown for specified metal/process/thickness UNVERIFIED. Marketing “pinhole-free” is not holiday-test acceptance or guaranteed zero defects.	UNVERIFIED; no actual alloy/insert seam cycle dataset located.
Thermal-sprayed oxide ceramic,wetted/exterior development fallback,NOT AlN chip substitute	BryCoat official page scraped typical0.003–0.030 inch (76.2–762 µm) and oxide choices; official-web lead only. Thus100/150 µm lie within a generic thickness range,50 µm is below that range,not proof of a selected process. Spray needs access before closure,finishing,porosity sealing and bondcoat control; line-of-sight misses hidden channels/seams. Do not assume ordinaryAlN survives a hot oxidative spray process unchanged.	Numerical exact-process dielectric strength UNVERIFIED; porosity,roughness and moisture can create weakpoints. Metallic bondcoats must never bridge intended electrical isolation.	UNVERIFIED; no selectedAl/Cu pressure-plate cold-cycle evidence.

Selection conclusion: zoned ceramic architecture is approved as a study requirement; no coating recipe or50/100/150 µm AlN production process is selected or qualified. An all-denseAlN coupon/foil bonded to the chip pedestal is a different construction,not demonstrated deposited coating; it would require Founder acceptance as satisfying the ruling plus explicit ceramic/bond/edge thickness and thermal budgets. Do not silently substitute bulk substrates,anodic chip faces or polymer TIM for the required chip dielectric.

D. Interface, corrosion and manufacturing hazards

Specify actualAl alloy/temper,Cu grade,insert construction and complete pressure boundary. A solidCu insert press-fit/brazed/welded intoAl can remain electrically conductive toAl even if both coolant-exposed surfaces are coated. Coolant barrier is not electrical isolation of theCu/Al joint. If dissimilar-metal electrical separation is required,define a separate insulating joint/load path with quantified mechanical/thermal/leak effects. A continuous intact barrier can interrupt electrolyte access; pinholes,cracks,weld heat-affected zones and joint edges restore a galvanic path. Coatings protect only while intact.
Map zone transitions and mask connector contacts,grounding/bonding points,seal lands,O-ring grooves,weld/braze/bond surfaces,threads and precision datums deliberately. Masked bare pressure-seal surfaces must not become unprotected wetted metal. Coordinate a joining/coating sequence that gives continuous protection without welding through ceramic or leaving an inaccessible uncoated seam. Coating before welding can be damaged; coating after closure needs demonstrated precursor/flow/spray access and cleaning. No final sequence is yet verified.
CeramicAlN/Al/Cu CTE mismatch,sharp insert corners,roughness,pores,particles,clamp contact and edge fields can initiate cracks or dielectric failure. Adhesion interlayers and stress control alter thermal impedance and growth; document all layers. HotCVD/ALD processing may change alloy strength,warp plates or damage joint/seals; not every advertised≤380°C process is acceptable to the pressure vessel. Coupon substrate/process thermal budget must be agreed before procurement.
Do not default wetted ceramic toAlN. Published liquid-cooling AlN study identifies aqueous working-fluid hydrolysis as a barrier toAlN use; this is a publisher-web lead,not validated coolant compatibility. Powder kinetics and bulk results do not quantify our deposited film in glycol/inhibitors. Actual water content,pH,conductivity,halides,inhibitor depletion,oxygen,flow/erosion andtemperature can matter. ChipAlN is nominally dry but still needs humidity/condensation recovery testing. OxideALD/PEO also needs actual coolant soak validation,not generic corrosion marketing.
Exterior ceramic is dielectric/corrosion protection,not dehumidification or guaranteed hydrophobicity. Inspect abrasion at guides/clamps,masked chassis contacts and field creepage; any coating-dependent insulation requires controlled minimum coverage and post-service inspection. Keep nitrogen purge/dew-point controls underICD§2.2 and warm admission underArticle5.

E. Qualification gates — provisional engineering targets, NOT ICD-approved requirements

Inputs still required before acceptance can be fixed: coolant product/composition/water fraction/inhibitors,cleanliness/pH/conductivity limits,min/max temperature,flow,working pressure,proof pressure and transients; actual module contact maps/powers,clamp force/TIM/stop geometry; actual maximum service/hot-end (not merely+7°C admission),temperature ramps,dwells,lifetime cycles; isolation topology/operating voltages/surge waveform/grounding and applicable safety standard. The44–59.5 V module window is not itself a certified dielectric test specification. Article5 supplies+5°C floor/+7°C release,−20°C routine and−40°C qualification mode; it does not define the highest plate temperature or cycle count.

Proposed screening population: at least5 representative coupons per process/substrate/thickness/zone combination and3 complete sealed plates, plus uncoated controls and intentional holiday/seam-defect controls. Include actualAl/Cu insert joints,edge radii,coating transitions,masked seals,ports,welds/heat-affected zones and finishing; flatAl witness coupons alone cannot qualify the assembly. Counts below are provisional development targets,not statistical reliability proof or a safety certification.

Dimensional gate: map minimum/maximum layer thickness,flatness,roughness andcoverage atface/edges/channels/seams; cross-section witnesses. Require every finished part inside the selected thickness/datum caps and pressure design minima. Verify89/98 mm only after actual structure+seals fit the residual budget. Dimensional acceptance outside those stated caps requires revised stack/pitch,not waiver by prose.
Electrical gate,before and after each exposure: wet/dry insulation resistance and hipot between conductive plate/insert/coolant and any isolated contacts,including seam/edges. Propose test levels defined from the approved maximum voltage+transient envelope with a documented engineering margin; actual voltage,waveform,duration,IR minimum/leakage limit and margin TBD after isolation/safety review. Do not apply an arbitrarykilovolt target as an asserted safety rating or multiply bulkkV/mm byfilm thickness to approve it. Require zero flashover/puncture atapproved levels and no worsening beyond approved leakage/IR limits. Performpinhole/holiday testing over100% accessible wetted barrier with detection voltage/method validated against intentional defects so inspection does not puncture good thin films; destructive witnesses assess inaccessible surfaces.
Coolant gate: provisional1,000 h circulated soak atactual hot-end/pressure/chemistry pluscold soaks,with intermediary100/250/500/1,000 h inspections andelectrical repeats; includeend-of-life inhibitor chemistry and controlled defect witnesses. Monitor galvanic current,pH/conductivity,dissolvedAl/Cu,gas/particles,coating mass/thickness andpit depth. No through-barrier attack,blistering/delamination orleak is a provisional pass expectation; numerical acceptable corrosion rate/ion release and galvanic-current limits TBD from wall-life/coolant budgets. Do not claim a1,000 h pass equals product lifetime.
Thermomechanical gate: provisional 100-cycle feasibility screen, then 1,000 routine cycles −20°C ↔ actual maximum service/hot-end and 500 qualification cycles −40°C ↔ that same hot-end on the actual clamped assembly. Traverse and verify +5/+7°C admission control during representative restart sequences; a +7°C-only hot endpoint is insufficient unless it actually bounds service. Ramps/dwells are TBD from actual thermal inertia and the worst credible warm-admission profile; dwell until representative interfaces stabilize. Repeat wet/dry electrical, pinhole, adhesion/tensile tests at checkpoints. Provisional retention ≥80% of initial adhesion/tensile strength with zero interfacial flaking, cracking through the barrier or leaks; absolute MPa minimum is TBD from stress/clamp/CTE analysis. Cycle counts/retention targets are proposals, not ICD approvals or supplier-qualified results.
Humidity/service gate: provisional 100 dock/open/reclose dry-purge recovery cycles and 100 controlled humidity/condensation excursions followed by approved dry recovery; atmosphere, RH, dwell and dew-point interlocks are TBD to actual facility exposure. No payload admission while moisture controls fail. Require no corrosion path, flashover/delamination and passing post-recovery IR/hipot. Exterior ceramic does not authorize wet operation or removal of nitrogen safeguards; evaluate purge failures safely on instrumented test hardware.
Thermal gate: measure through-plane thermal impedance of the actual Al/Cu/coating/TIM/contact stack at specified clamp load and compression stop, at −40/−20/+7°C and the actual hot-end as applicable; separate coating/bond resistance from TIM and spreading using controls. Verify CPU/OAM maps at ICD envelopes and actual transients. Provisional ≤10% increase in complete pickup thermal impedance after exposure, plus all actual Tcase/junction limits met; absolute K/W allowance and temperature margins are TBD pending selected modules, coolant and contacts. Measure k_z(T) rather than borrowing bulk 170 W/(m·K).
Pressure/leak/hydraulic gate: inspect finished channel reduction and flush cleanliness; map flow/pressure drop with actual cold-coolant viscosity. Pressure/leak test before/after clamp and thermal cycles and at service limits; proof/burst pressures, allowable leak rate, minimum flow and maximum Δp are TBD by pressure-vessel/system requirements, not invented here. Recheck sealed-joint stresses and trapped-coolant warm expansion/relief. Require pressure integrity and flow delivery with no coating debris/obstruction; no wall-thinning credit without analysis/test.

Qualification means ALL gates pass for the actual manufactured,joined,sealed,clamped assembly and its controlled process—not a supplier claim,a bulk material property,a−40°C service label or one flatcoupon. Failure in any requiredzone leaves the plate unqualified. Continue no schematic/PCB implementation untilgovernance gates close.

R-CL8A verification and scope status

Completed: ruling recorded; entire ICD reread; bounded 50/100/150 µm external/datum/channel budgets and conditional 89/98 mm conclusion; derived CPU/OAM heat-flux/ΔT arithmetic; validated bulk and actual thin-film PDF citations; bounded official supplier/paper survey with process-specific unknowns; interface hazards and provisional qualification plan. The calculation tool checked fluxes, film R'', ΔT and 4 mm residuals. Read-only document readback confirmed the complete added section, its three tables and every new inline citation marker; the additive patch did not replace prior revisions. The read tool truncates the long pre-existing archive in its displayed response, so this is not a new line-by-line re-review of the entire historical archive.

Blocked/unestablished: actual 4 mm reserve, finished module/contact datums, qualified 50–150 µm AlN on Al/Cu, k_z(T), wet/dry dielectric limits, actual coolant/pressure/hot-end, Cu/Al joint isolation and complete cold-cycle performance. Unattempted by design: coating procurement/fabrication, FEA/hardware qualification, schematic/PCB mutations and any ICD amendment. No changed nets exist for ERC/DRC, and existing electrical issues are unaffected. This addendum adds tables only within R-CL8A; the original five Rev C request tables below remain unchanged. Earlier mass tables exclude the newly required ceramic, bond layers and finishing changes; coating area/density and actual material removal are not established, so those masses remain uncoated comparison references, not updated coated-assembly weights.

Rev C — stack evidence, paired rows and facility comparison

Basis: complete Rev C request, read in full; governing ICD Rev A. This is documentation/research/calculation only. All Rev B results and the archived Rev A remain below, unchanged. Rev C changes its own service-floor convention to two opposing rows sharing one lane, whole units removed INTO that lane, no additional rear apron; this does not retroactively change Rev B's one-sided-row/ring areas. Study revision C is not an ICD revision, architecture approval, qualified product comparison or permission for component placement/routing under Article10.

Status: verified means bounded evidence supports the stated fact, not the whole assembly; calculated means arithmetic on declared inputs; assumed/unverified means no qualified complete design or published dimensional proof. N/D means not determinable from available evidence, never zero. The completed bounded OAM/CPU and GB200 research handoffs were retrieved and reused, rather than repeating a broad vendor survey. There are exactly five Rev C tables below, one per numbered request, followed by new assumptions/unresolved items.

1. Complete OAM-face stack height

Define carrier OAM-facing surface as z=0, module underside z_b, module PCB thickness t_M, component thermal-contact height a_i above module top, compressed EXTERNAL TIM b_i and pedestal projection d_i from the common plate inner plane toward pickup i. Then, as a datum equation rather than a vendor dimension:

h_O=z_b+t_M+\max_i(a_i+b_i+d_i)

Other tall components/clearances also bound the plate position; worst-case bow, coplanarity, attach, TIM compression, pedestal/plate flatness and thermal deformation must be included. The package/HBM branches are parallel pickup heights, not serial layers to add. If a_i already includes attach/BGA, substrate, interposer, chip/HBM and lid/internal TIM, do not add those again. External TIM and pedestal must not be hidden or omitted.

OAM r2.0 lists 5±0.15 and8±0.15 mm connector stacks. The5 mm option uses218910-1115 on module and baseboard;8 mm uses218916-1115 on module with218910-1115 on baseboard. Recommended bottom-stiffener separations likewise are5±0.15/8±0.15 mm. These are not connector+stiffener spacers to add in series. Exact production-drawing dimension endpoints remain unresolved: if s is carrier-surface-to-module-underside, use z_b=s and add PCB once; if s terminates at module top, use z_b+t_M=s and do not add PCB; housing/BGA datums require their own conversion first.

For the table only, X_A/X_G denote the complete remaining module PCB+thermal-pickup+TIM+pedestal stack, with appropriate tolerances, under the unverified underside-datum hypothesis. Neither X is numerically established. Source: OCP r2.0 v1.0. No numerical best verified complete hO was found for either shipping accelerator; a form-factor confirmation is not full mechanical data.

With hC35 mm, unchanged3 mm carrier,two6 mm plates,4 mm local structure and9 mm guide/sleeve pitch additions:

T=54+h_O;\quad p=63+h_O;\quad N=\lfloor340/(63+h_O)\rfloor;\quad K=RN;\quad P_{nom/peak}=(RN-1)(1.5/1.8)+0.320

E-compatible H230,rung254 gives R6,total1,764 mm with200 mm headers+40 mm walls. D-only H201,rung225 gives R7,total1,815 mm; D-only does not satisfy an E-compatible envelope. Keep250 mm depth and the90 mm utility reservation,50 mm dock,+20 mm service depth. None is a proven converter/hardware fit.

Table

Accelerator / option	Verified bounded evidence	Complete hO and status	Recomputed pitch /slots per rung	Physical /active per unit; nominal /peak	Four/rung reachable?
AMD Instinct MI300X,5 mm option	Official datasheet confirms OAM module.	5+X_A only under datum hypothesis; unverified/N/D; shipped5 mm compatibility not verified	p=68+X_A; N=floor[340/(68+X_A)], numerical N/D	E:6N physical /6N−1 active; D:7N /7N−1; numerical power N/D	Not proven; needs complete worst-case hO≤22 mm
MI300X,8 mm option	Same verified OAM identity; exact shipped connector/r2.0 mechanical compatibility unresolved	8+X_A conditionally; unverified/N/D	p=71+X_A; N=floor[340/(71+X_A)], numerical N/D	Same RN−1 formulas; numerical population/power N/D	Not proven
Intel Gaudi3 HL-325L,5 mm option	Official brief confirms OCP Accelerator Module V2.0 compliant.	5+X_G conditionally; hypothetical connector configuration, unverified/N/D	p=68+X_G; N=floor[340/(68+X_G)], numerical N/D	Same RN−1 formulas; numerical population/power N/D	Not proven; cannot convert a shipping8 mm assembly by arithmetic
HL-325L,8 mm option	Same verified OAMv2 identity; OEM lead below is incomplete	8+X_G conditionally; unverified/N/D	p=71+X_G; N=floor[340/(71+X_G)], numerical N/D	Same RN−1 formulas; numerical population/power N/D	Not proven
Case6 inherited allowance,not a module measurement	hC35+hO35 assumed	hO35,T89 mm	p98 mm;3/rung	E18 /17,25.82 /30.92 kW; D21 /20,30.32 /36.32 kW	No under this allocation
Threshold sensitivity,NOT best verified hardware	hC35+hO22 hypothetical	hO22,T76 mm	p85 mm;4/rung,zero lateral reserve	E24 /23,34.82 /41.72 kW; D28 /27,40.82 /48.92 kW	Conditional equality only; no shipping-module proof

Sources: AMD MI300X datasheet, Intel HL-325L brief. Missing for both: released module PCB/package/HBM contact-map dimensions, selected mated configuration, compressed external TIM,pedestal and full tolerances. An OEM Supermicro Gaudi3 lead reports4.7 mm PCB,8 mm connector and8±0.15 mm bottom stiffener including Mylar; these are web-derived,unverified as Intel dimensional evidence and not a complete hO. Its contradictory power descriptions further prevent treating it as the released mechanical authority.

Four/rung requires hO≤22 mm; three/rung requires hO≤50.333 mm. Under the underside hypothesis, nominal5/8 mm options leave17/14 mm for the entire remainder, NOT just chip height; using connector upper bounds5.15/8.15 leaves16.85/13.85 mm before all remaining worst-case terms. An illustrative4.7 mm PCB would leave12.3/9.3 mm nominal or12.15/9.15 mm at those connector upper bounds; this is a conditional budget, not Gaudi fit evidence. A3 mm connector reduction is valid only if the module/stiffener and all other geometry permit that option. Keep the earlier dual-sided bolt/bolster,pad/via/backdrill,metal clamp-load and Gen5 channel blockers; no confirmed routed length follows from overlap.

2. Two facing rows with a shared service lane and whole-unit N+1

All installed units are included at their one-sled-spare operating envelope; an extra installed unit provides aggregate whole-unit N+1 capacity, not a silently compute-off tower. Each unit also retains one installed compute-off sled spare. The capacity criterion is after loss of any one equal-rated unit:

n=\lceil900/P_{nom}\rceil+1;\quad(n-1)P_{nom}\ge900

n_1=\lceil n/2\rceil;\quad n_2=\lfloor n/2\rfloor;\quad L=n_1W+(n_1-1)0.020+1.000;\quad A=L(2D+\ell)

Use20 mm inter-unit gaps within each row and0.5 m at each row end; allocate the rectangle using the longer row even if the other has one fewer unit. Manual lane ℓ1.20 m; rail lane ℓ1.30 m. No rear apron is added in Rev C; utilities must be disconnected safely before whole-tower removal INTO the lane. Only selected sled/tower service occurs at a time; no facing simultaneous extraction is assumed.

Table

Requested unit case /status	Unit physical /active; nominal /peak kW	Installed units; row split; installed sled physical /active /spares	System nominal /peak; after one unit loss nominal	Longer-row L	Shared lane manual /rail	Total paired-row area manual /rail	Tower-handling clearance check
Standard side-by-side400×360,t20	16 /15;22.82 /27.32	41;21+20;656 /615 /41	935.62 /1,120.12;912.80 kW	9.800 m	1.20 /1.30 m	18.816 /19.796 m²	D0.36+person0.75=1.11 m; residual0.09 /0.19 m before cart/fixtures; not qualified
Case6 E-compatible35+35,assumed	18 /17;25.82 /30.92	36;18+18;648 /612 /36	929.52 /1,113.12;903.70 kW	8.540 m	1.20 /1.30 m	16.3968 /17.2508 m²	Same0.09 /0.19 m residual; whole-tower mass/support unresolved
Case6 E,35+best verified hO	N/D; no verified complete hO	N/D, not omitted or substituted	N/D	N/D	Requested1.20 /1.30 m	N/D	Geometric D remains0.36 m provisionally; fit still unproven
Thin side-by-side423.5×340,t10,REFERENCE ONLY; ≤25 mm active stack unverified	24 /23;34.82 /41.72	27;14+13;648 /621 /27	940.14 /1,126.44;905.32 kW	7.189 m	1.20 /1.30 m	13.5153 /14.2342 m²	D0.34+person0.75=1.09 m; residual0.11 /0.21 m before cart/fixtures; not qualified
EXTRA sensitivity: Case6 E,35+22,NOT verified hO	24 /23;34.82 /41.72	27;14+13;648 /621 /27	940.14 /1,126.44;905.32 kW	6.860 m	1.20 /1.30 m	13.1712 /13.8572 m²	Same0.36 m tower-depth check; equality pitch and module compatibility unresolved

Counts and powers are calculated, all hardware fits conditional. Whole-unit N+1 here is a power/population property, not an end-to-end availability percentage or application fault-tolerance guarantee. The extra unit adds another sled spare; a spare tower is not a second spare in every rung. Redundancy of plant,cooling,fabric,governance and shared service equipment is unresolved. The1.20 m manual lane normally reserves0.45 m fixture+0.75 m person;1.30 m rail reserves0.55 m robot/track+0.75 m person. Whole-tower0.36 m depth replacing the retracted tool nominally fits the scalar budget, but the small residual is not proof of cart width,turning,foot reaction,stability,guide stroke or egress. Park/retract hood/rail equipment and serialize tower service; do not add its occupied band again or pretend the person can safely stand in a robot swept zone. Rear utility access must be front-serviceable or obtained by safely extracting the isolated tower; otherwise the no-rear-apron premise fails.

3. GB200 NVL72 baseline, limited-boundary ratios and two levels

Requested verified ALL-IN900 kW hall area is N/D. No complete published layout with all support allocations and clearances was established. The following is a transparent partial example, not an all-in density claim. Use one consistent OEM configuration: HPE GB200 NVL72 QuickSpecs and HPE product page, not HPE mass combined with a different DGX120 kW rating. These HPE numbers were checked against official web material in the research handoff; PDF-citation verification and installation-specific qualification remain unverified.

HPE publishes132 kW nominal per rack,approximately1,471.9 kg fully populated/coolant-filled with PGW,and0.600 m width×1.068 m length footprint. Heading/order ambiguity in its dimensions requires installation-drawing confirmation;0.600 is taken along the row and1.068 along service depth here. Six,max eight33 kW shelves are integrated, as are busbars,ToR management and manifolds: do NOT add a duplicate shelf rack for those shelves. Published132=115 kW liquid+17 kW residual air per rack; seven racks therefore imply805 kW liquid+119 kW air at that stated split. HPE lists1.3 MW in-row CDU support for at most eight racks; this alone does not establish required flow,approach-temperature duty or redundancy. Its peak wording says192 kW and1.5×TDP,which would be198 kW at132 kW: conflict unresolved; neither value is used for electrical sizing.

The original Vertiv XDU1350 installation guide Rev C research handoff reports0.900×1.242 m footprint,684 kg dry/875 kg operating; source-derived but not Flux-PDF-verified in this synthesis. Do not substitute XDU1350B or its mass. The1.20 m front and1.20 m rear aisles below are UNVERIFIED ANALYST ALLOCATIONS, not HPE/Vertiv minimum service clearances; no end/cross-aisle or further handling margin is in this GB subtotal.

n_{GB}=\lceil900/132\rceil=7;\quad A_{rack+aisles}=7(0.600)(1.068+2.400)=14.5656\ m^2

A_{CDU}=0.900(1.242+2.400)=3.2778\ m^2;\quad G=17.8434\ m^2

Ratio definition is ColdNative paired-row area / G, not its inverse; smaller than1 means smaller ONLY at the expressly incomplete numerator/denominator boundaries. Both sides require further support area. ColdNative includes one whole-unit spare while the primary seven-rack GB baseline is not compute-rack N+1; an eight-rack sensitivity makes that asymmetry visible. Equal kW is not equal compute throughput,AC input,PUE or workload performance.

Two-level ColdNative figures are idealized geometry only: stack two2.0 m units vertically at each footprint position, plus unallocated supporting structure and overhead lifting/gantry clearance. Use F=ceil(n/2) ground positions,split ceil(F/2)/floor(F/2) between facing rows; retain full shared lane,end allowances and gaps. Thus:

L_2=\lceil F/2\rceil W+(\lceil F/2\rceil-1)0.020+1.000;\quad A_2=L_2(2D+\ell)

Never divide an already aisle-inclusive area by2. For odd n,one upper location is empty; no extra unit,power or mass is invented. Upper-tier lift/handling range must cover another2 m plus structure,with an overhead tower-removal route,not just the old1.4 m sled lift. No load-bearing mezzanine or two-story floor is approved.

Table

Baseline /ColdNative case	Power/population basis	Partial floor area manual /rail,or stated GB allocation	Two-level ground positions;paired split;L2	Two-level partial area manual /rail	Single-level ratio to G manual /rail	Two-level ratio to G manual /rail	Mass /kg per m² boundary and qualification
Requested ALL-IN GB200 hall baseline	≥900 kW	N/D,missing support/clearance allocation	Not applicable	Not applicable	All-in ratios N/D	All-in ratios N/D	All-in operating mass and density N/D
HPE seven racks +one original XDU1350,worked PARTIAL subtotal	7×132=924 kW;lose one→792,not N+1	G17.8434 m²,including assumed front/rear aisles; bare rack sum4.4856 m²	Not applicable	Not applicable	Reference1.000	Not applicable	Rack1,471.9 kg wet/2,297 kg/m² own footprint; subtotal11,178.3 kg wet/626.5 kg/m²; web-derived inputs,assumed aisles
HPE seven racks +two original CDUs,sensitivity only	924 kW;CDU hydraulic/redudancy topology unverified	21.1212 m²	Not applicable	Not applicable	Not the chosen G	Not applicable	12,053.3 kg wet/570.7 kg/m² limited boundary; no qualified redundancy claim
HPE eight racks +one CDU,compute-rack N+1 sensitivity	1,056 kW;lose one→924	19.9242 m²,same assumed aisle model	Not applicable	Not applicable	Not the chosen G;compare like N+1 separately	Not applicable	CDU support says max eight racks;full duty/utility redundancy still unresolved
Standard ColdNative,item2	41 units,935.62 kW	18.816 /19.796 m²	21;11+10;5.600 m	10.752 /11.312 m²	1.055 /1.109	0.603 /0.634	ColdNative dry unit mass excludes coolant,rail/gantry and external support;not a wet all-in comparison
Case6 E35+35,assumed	36 units,929.52 kW	16.3968 /17.2508 m²	18;9+9;4.760 m	9.1392 /9.6152 m²	0.919 /0.967	0.512 /0.539	Plate material masses in item4;two levels do not halve installed mass
Case6 E35+best verified hO	N/D	N/D	N/D	N/D	N/D	N/D	No complete hO;cannot claim a best-verified density result
Thin side-by-side,t10,reference conditional	27 units,940.14 kW	13.5153 /14.2342 m²	14;7+7;4.0845 m	7.6789 /8.0873 m²	0.757 /0.798	0.430 /0.453	Thin CPU/OAM stacks and10 mm walls remain unqualified
EXTRA Case6 E35+22 hypothetical	27 units,940.14 kW	13.1712 /13.8572 m²	14;7+7;3.920 m	7.5264 /7.9184 m²	0.738 /0.777	0.422 /0.444	Not verified hardware;not a replacement for the N/D row

Operating-mass arithmetic consistently uses the rounded1,471.9 kg/rack:7 racks=10,303.3 kg,+875=11,178.3 kg; the research handoff's11,178.5 kg rounded subtotal differed by0.2 kg and is not propagated. This is far below the real mass uncertainty. Integrated power shelves are included in rack dimensions/mass;not free space and not duplicated off-rack. A300 mm optional rack-extension sensitivity would add1.2600 m² to seven rack service allocation before CDU/support additions.

NVIDIA DGX GB200 hardware guide separately reports about120 kW/rack;eight racks give960 kW but exact mass was not verified,so do not combine that power with HPE mass. NVIDIA SuperPOD architecture includes networking,management and storage beyond compute. Vertiv Reference Design020 illustrates substantial support racks,CDUs,PDUs and residual-air cooling in a different1,152 kW POD;it must not be linearly scaled and called a verified900 kW installation.

For either platform the missing all-in support term includes any separately allocated fabric/network/storage/management racks,residual-air equipment where applicable,pumps/CDUs and their true service clearances,hall PDUs/disconnects,plumbing/gas equipment,cross aisles,egress and handling. Boundary choice must explicitly include or exclude UPS,switchgear and external heat rejection on BOTH sides. ColdNative's conduction-only ICD is unchanged;GB200's residual-air equipment is a comparison-system requirement,not a new ColdNative air-cooling rule. All-in area is A_partial+A_support,with A_support presently unknown;no verified all-in advantage is established by the ratios.

4. Case6 plate materials and unit dry mass

Use the SAME E-compatible Case6 unit for all materials:400×360 mm outside,t20,six254 mm rungs,three98 mm positions per rung,18 physically installed sleds/17 active,35+35 assumed stacks and25.82/30.92 kW. Height savings do not reduce the2 m enclosure/hardware estimate. Both plates230×250×6 mm;20% channel/pocket void;densities2,700 Al/8,960 Cu kg/m³ are inherited study assumptions,not material-strength evidence. Keep non-plate sled remainder 6.000−2.6568=3.3432 kg, with no invented module/converter/stiffener reduction.

Hybrid allocation assumes one50×50×6 mm CPU insert in its facing plate and one80×80×6 mm OAM insert in the opposite plate,replacing aluminum through full6 mm thickness,not adding copper on top. Their number,dimensions and die coverage are analyst assumptions,not actual MI300X/Gaudi/CPU thermal maps. Inserts are solid;20% void applies only to remaining aluminum. This explicitly fills the otherwise modeled20% void under those regions;it is not a claim that coolant channels can intersect solid inserts or that the interface is pressure-qualified.

V=2(0.230)(0.250)(0.006)=0.000690\ m^3;\quad V_i=(0.050^2+0.080^2)(0.006)=0.0000534\ m^3

m_{Al}=0.8V\rho_{Al};\quad m_{Cu}=0.8V\rho_{Cu};\quad m_H=0.8(V-V_i)\rho_{Al}+V_i\rho_{Cu}

M=18(3.3432+m_{plates}+1.2)+51+6(4)+2.48832;\quad\sigma=M/0.144

Fixed stationary allocation is 75 kg=51+six×4,not the old four-rung67 kg. Retain2.48832 kg insulation bounding allocation and 1.2 kg permanent sleeve hardware per physical slot as a conservative estimating bound,not a measured mass: analogous234 mm-high,91 mm-clear-width sleeve with98×241 mm cap at the earlier1 mm-sheet model gives about0.883 kg including0.20 kg hardware;retain1.2 kg to avoid unsupported lightening,subject to actual seals/stiffening. This is not a certified upper bound on a released assembly. Coolant,external head/rail/gas equipment and stack-support gantry remain excluded. Keep ±30% planning uncertainty;rounded results are arithmetic,not measured precision.

Table

Case6 material variant	Dry plate pair /removed sled mass	Unit dry mass,all18 slots	Own-footprint dry kg/m²	Ratio to HPE rack-own-footprint wet2,297 kg/m²	Qualification /comparison boundary
Aluminum plates	1.49040 /4.83360 kg	186.1 kg	1,292	0.563	Calculated on assumptions;dry versus wet,not a hall-floor or structural advantage claim
Copper plates,same void geometry	4.94592 /8.28912 kg	248.3 kg	1,724	0.751	Same boundary mismatch;handling and plate/load qualification open
Aluminum with two solid Cu replacement inserts,dimensions above	1.85352 /5.19672 kg	192.6 kg	1,338	0.582	Extra0.36312 kg/sled overAl includes solid-insert void convention;thermal/galvanic/seal durability unverified
HPE rack reference,item3	Coolant-filled rack≈1,471.9 kg;not a ColdNative sled	Not a ColdNative unit	2,297 wet kg/m²,rack footprint0.6408 m²	1.000 reference	Official-web-derived mass/dimensions,PDF/installation qualification unverified;partial-room626.5 kg/m² is a DIFFERENT denominator and not used above

No claim of thermal equivalence among Al,Cu and inserts follows from mass. Validate actual die/HBM pickup area,spreading resistance,interface contact,CTE,galvanic isolation,pressure integrity and coolant chemistry. Two-level support reactions include BOTH units plus coolant,frame and machinery;own-footprint kg/m² is not support-foot pressure,slab rating or an approved two-story loading plan. Wet mass equality and full hall averages cannot be supplied without coolant inventory and common equipment/area boundaries.

5. COM-HPC Server D/E CPU survey — information only

No exact standard Server D/E ordering variant with BOTH a Gen5 host link and the ICD's−40°C-minimum Tcase guarantee was verified in this bounded survey. This is not proof that none exists worldwide. Connector/form-factor Gen5 capability is not an actual CPU-module interface claim,and a processor's cold rating,storage temperature or custom-temperature marketing is not a qualified module ordering variant. No selection/substitution is made. Complete populated mated stack remains unverified for every listed family;the35 mm CPU allocation remains a study assumption.

Table

Family /exact variants examined	Server outline /interface status	Host PCIe /memory	Temperature evidence	Complete populated carrier-to-plate stack	Result against Gen5+cold standardD/E requirement
conga-HPC/sILH, 050900 (D-2796TE), 050901 (D-2775TE), 050902 (D-2752TER), 050995 (D-2796TE), 050996 (D-2775TE)	Server D family; optional E family exists, but do not infer exact E SKU from these strings	32 Gen4 +16 Gen3, not Gen5; DDR4 ECC/non-ECC; PN050902:2667 MT/s, others:2933 MT/s.	−40°C CPU Tcase minima; maxima83/82/86/83/82°C respectively.	Unverified; memory class/population changes full height; up to512 GB LRDIMM /256 GB RDIMM /128 GB UDIMM per manual.	Verified exclusion: Gen4 despite cold evidence
conga-sdIL /COMh-sdIL E2 family; exact E2 ordering strings not validated	D small120×160 mm, not standard D/E	16 Gen4 +16 Gen3; soldered DDR4 ECC up to64 GB.	E2 section:−40…+85°C. Technical table:−40…+80°C operating. Ambient/Tcase and maximum-rating discrepancy unresolved	Unverified; memory-down does not establish full height	Excluded: nonstandard outline and Gen4; temperature conflict remains open
Advantech SOM-E781S64-U3A1, SOM-E781S16-U4A1	Size E-class candidate, but proprietary COM-HPC pinout.	72 Gen5 +7 Gen3, DDR5 RDIMM4800 MT/s.	Exact variants: 0…50°C operating, not extended cold. Ambient/Tcase definition unresolved	Unverified;7.8–8.6 mm IHS-to-MB validated range is NOT complete mated populated height.	Excluded: no cold rating for these exact variants and not a standard interchangeable pinout
Portwell PCOM-B800GT /PCOM-B801GT, secondary web exclusions	B800GT Size E /B801GT Size D marketed	B800GT Gen4, DDR4 DIMMs; B801GT Gen3, per official web survey	B800GT−40…85°C marketed by CPU SKU; exact cold ordering variant/Tcase not verified	Unverified	Official-web leads only, PDF-unverified; excluded on Gen4/3, no qualifying part claimed

Manuals: conga-sILH HEIH, conga-sdIL, SOM-E781; Portwell official web source. Newer Gen5 Client A/C search leads are not Server D/E substitutions. Require a vendor-written exact Gen5 standard Server D/E SKU, −40°C Tcase guarantee, qualified memory configuration, pin map and complete populated mated STEP before closing this item.

Rev C — new assumptions, unresolved items and verification

Item1 completed as an evidence-status study; numerical complete hO failed to establish. AMD MI300X and Intel HL-325L are shipping-family examples, not selected modules. Obtain released module STEP/mechanical thermal guides, PCB/package/HBM contact map, connector/stiffener BOM and Molex2189101115-SD/2189161115-SD drawing datums. The family PDF lookup remained pending in the research handoff, not absent. No verified best hO, guaranteed22 mm fit or automatic5 mm shipping variant is claimed.
Datum/tolerance closure: define hO and hC to inner plate planes, include compressed TIM/pedestals and clearance, avoid adding one PCB/package/internal TIM twice, and correlate connector/standoff constraints in one assembled stack. Actual bolster/screw/pad/via escape interference may force offsets and larger H/depth; zero axial reserve remains unqualified.
Item2 completed conditional paired-row calculations. Whole-unit N+1 is satisfied numerically with41/36/27 units for the standard/35+35/thin bases. The best verified hO case is explicit N/D; the22 mm row is an additional hypothesis, not a substitution. Shared-lane head parking, tower cart/rail guidance, stability, utility isolation and rear serviceability need an installation drawing and service validation. Safety/egress or ergonomic qualification is not established.
Item3 all-in baseline failed to establish; partial comparison completed. HPE132 kW and wet mass stay together; DGX120 kW is separate. Official web facts without validated PDF citations remain identified as web-derived/unverified in this synthesis. The1.2+1.2 m GB aisles and CDU service allocation are analyst assumptions. Resolve HPE dimension-heading and192/198 kW peak conflicts, CDU duty/flow, redundancy, network/storage/rack extensions, residual119 kW air cooling and all support areas. No verified all-in advantage ratio is available.
Two levels are unqualified geometry: two2 m unit stacks need more than4 m clear height once frame/lifting head/utilities are included. Overhead gantry, tower-removal path, upper-tier leak containment, anchoring, foundation loads and extended sled lift are unresolved; reported areas exclude all those increments. Shared lanes are not halved and installed mass/power do not decrease.
Item4 completed mass arithmetic, not thermal/mechanical design. Three plate variants use18 sleds, six rungs and unchanged non-plate remainder. Retained1.2 kg sleeve allocation and51 kg shared hardware are planning allowances, not qualified maxima;18/21+port fabric and larger manifolds may increase real weights and320 W shared power. Insert geometry, voids, joined pressure boundary, corrosion and thermal performance are unverified. No Cu-on-top mass double counting or removed rung fixed hardware occurs.
Item5 survey completed; qualifying exact CPU failed to establish. Gen4 cold SKUs, Gen5 commercial/proprietary variants and nonstandard D-small outlines are not approved substitutes. New vendor information may change the survey; no schematic selection was attempted.
Verification: scripted recalculation checked unit ceiling counts, N+1 after loss, RN−1 active counts, unequal row lengths, shared lanes, two-level integer positions, partial GB areas/mass denominators, all ratios and three plate mass variants. Read-only project-file readback checks heading/table persistence and inline citations; earlier Rev B/Rev A are preserved. No electrical/PCB mutation was attempted or performed, so there are no changed nets to review by ERC/DRC. This does not clear existing ERC issues or Article10 architecture/layout approval gates.

Rev B — rectangular variants

Basis: Rev B request. This is documentation-only. No schematic, PCB, footprint, component placement or routing change was made. Study revision B is not an ICD revision. Rev A remains below as the inherited reference; the following overrides its square-footprint assumption and explicitly identifies the service additions.

The governing footprint limit is 0.144 m², not 380 mm on each side. The requested 424 × 340 mm = 0.14416 m², exceeding the limit by 0.00016 m² = 160 mm² = 0.1111%. Its rows are retained as requested, marked noncompliant. A 423.5 × 340 mm = 0.14399 m² nominal rectangle corrects this without changing the unsealed lateral counts. The other requested rectangle, 450 × 320 mm, is exactly 0.144 m² nominal. Production dimensional tolerances must keep the maximum area inside the limit; an exact nominal-area result has no positive tolerance allowance.

The archived Rev A 380 × 380 mm square is itself 0.1444 m², 400 mm² above the newly stated area limit; it is retained only as historical comparison, not as a currently area-compliant envelope.

Calculation basis used in all case tables

W is the outside slot-bank dimension; D is the outside withdrawal dimension; t is wall thickness per side. Retain 20 mm total slot-bank end structure, four rungs, one installed compute-off spare per unit, 1.5/1.8 kW input envelopes per active sled, and the requested 320 W shared/spare allowance. “Unit kW” below includes that allowance and is nominal/peak. For N slots per rung:

B=W-2t-20,\quad N=\lfloor B/p\rfloor,\quad n_{\mathrm{active}}=4N-1

P_{\mathrm{nom}}=1.5(4N-1)+0.320,\quad P_{\mathrm{peak}}=1.8(4N-1)+0.320\quad[\mathrm{kW}]

All unit mass columns are dry, with every slot physically occupied, including the spare. Preserve Rev A's rounded 6.000 kg aluminum-plate sled and unchanged non-plate allocations. Identical copper plate geometry adds 6.15984 kg/sled, giving 12.15984 kg/sled. No unsubstantiated mass saving is credited for soldered memory or the 10 mm thinner package. Shared fixed mass stays 67 kg: 51 kg unit hardware plus four 4 kg rung allocations; new sleeve/seal hardware is additional. Preserve the insulation bounding-volume method at 2 m height:

m_{\mathrm{ins}}=40\{2WD-(W-2t)(D-2t)(2-2t)\},\quad W,D,t\text{ in metres}

M=4N(m_{\mathrm{sled}}+s)+67+m_{\mathrm{ins}},\qquad \sigma=M/(WD)

Here s is zero for unsealed comparison rows and 1.2 kg per physical slot for the new permanent service sleeves/doors. The service fixture outside the unit is excluded. Rev A's ±30% unit-mass planning uncertainty remains; one-decimal masses and integer kg/m² are calculation outputs, not measurement precision. Values in each “Al / Cu” cell correspond to aluminum or copper cold plates only; enclosure and service sleeves remain aluminum.

Case 1 — 424 × 340 mm, 20 mm walls, 65 mm pitch

Clear withdrawal depth is exactly 300 mm: 250 mm sled + 50 mm dock, with no unallocated axial space. Slot-bank width is 364 mm, or 363.5 mm after trimming W to 423.5 mm. These are packing-only rows, not live-service-ready units.

Table

Envelope / status	Slots/rung	Physical / active slots per unit	Unit kW nominal / peak	Unit dry mass Al / Cu	Unit kg/m² Al / Cu
Requested 424 × 340; area exceeds limit, no new sleeve	5	20 / 19	28.82 / 34.52	189.5 / 312.7 kg	1,315 / 2,169
Area-corrected 423.5 × 340; no new sleeve	5	20 / 19	28.82 / 34.52	189.5 / 312.7 kg	1,316 / 2,172

Insulation allocations are 2.50112 kg and 2.49928 kg respectively. The lateral bank has 39 mm left after five 65 mm positions in the requested rectangle. Full sealed-service results are in Case 4; the 20 mm added depth cannot be ignored.

Case 2 — 450 × 320 mm, 10 mm walls, 65 mm pitch

Clear withdrawal depth is again exactly 300 mm. Slot-bank width is 410 mm; six 65 mm positions occupy 390 mm, leaving 20 mm. Footprint is 0.144 m² nominal.

Table

Envelope / status	Slots/rung	Physical / active slots per unit	Unit kW nominal / peak	Unit dry mass Al / Cu	Unit kg/m² Al / Cu
Requested 450 × 320; no new sleeve	6	24 / 23	34.82 / 41.72	212.3 / 360.1 kg	1,474 / 2,501

Insulation allocation is 1.30320 kg. This row is laterally and axially nominally packable, but not a demonstrated coupling, tolerance, cold-service or thermal fit.

Case 3 — both rectangles at 55 mm pitch / 50 mm sled

Retain 3 mm carrier, 6 mm rear zone, two 6 mm plates and 4 mm local sealing/structure allowance. The complete active-side stack must therefore shrink from 35 mm to:

h_{\mathrm{active}}\le50-(3+6+12+4)=\mathbf{25\ mm}

This is the whole mated stack above the carrier surface, including carrier-to-module connector spacing, module PCB, populated memory, thermal pickup/TIM and any other protrusion inside that allocation—not the RAM PCB's height alone. The OAM stack also has to meet the same local thickness envelope. A lower CPU-memory stack does not prove a 50 mm full sled.

Table

Envelope / status, 55 mm pitch	Slots/rung	Physical / active slots per unit	Unit kW nominal / peak	Unit dry mass Al / Cu	Unit kg/m² Al / Cu
Requested 424 × 340, t=20; area exceeds limit, no sleeve	6	24 / 23	34.82 / 41.72	213.5 / 361.3 kg	1,481 / 2,507
Area-corrected 423.5 × 340, t=20; no sleeve	6	24 / 23	34.82 / 41.72	213.5 / 361.3 kg	1,483 / 2,509
Requested 450 × 320, t=10; no sleeve	7	28 / 27	40.82 / 48.92	236.3 / 408.8 kg	1,641 / 2,839

Module-family answer: low-profile/soldered-memory server-family options exist, but no complete ≤25 mm active stack satisfying all current ICD requirements has been verified. Manufacturer web material lists VLP-RDIMM support for conga-HPC/sILH, a standard Size D family with an optional Size E variant. VLP memory is still a DIMM; it is not proof of the complete mated height. The conga-sdIL product offers soldered ECC RAM and is marketed as Server Size D, but its manual calls it “Size D small,” 120 × 160 mm, with Server-D-aligned pinout/connector placement, not a standard 160 × 160 mm outline. It is a lead requiring explicit form-factor acceptance, not a silent D/E substitution. Both product lines' web material indicates PCIe Gen4, conflicting with the ICD's Gen5 host-link requirement. ADLINK COM-HPC-ALT was also examined: Size E with socketed RDIMMs, not a verified soldered-memory solution.

Evidence status: module dimensional and memory datasheet queries remain pending/unverified in this session. The candidate web findings above do not establish full-stack height, written −40°C Tcase qualification, mechanical interchangeability or Gen5 compliance. No module is selected, and the 55 mm rows remain conditional arithmetic. Published module-only heights or a VLP memory height must not be substituted for carrier-to-tallest-point stack height.

Case 4A — per-slot sealed service with dry-gas flush

Concept: permanently subdivide each rung into individual sealed slot sleeves, with a sealed rear dock/bulkhead around power, fabric, coolant and gas penetrations. A door alone on a common open rung volume would not isolate neighboring sleds. Each sleeve gets an individually removable gasketed front cap, isolated purge supply/vent valves and a dry removable service hood. The permanent subdivision is an explicit refinement of the Rev A rung compartments, not a claim that a shared rung door is sufficient.

Study dimensions, not vendor-certified parts:

Keep the original 5 mm guide/assembly allowance; add 4 mm pitch per slot, covering two independent 1 mm sleeve side sheets plus 2 mm additional service/seal alignment allowance. Effective pitch becomes 69 mm for the 60 mm sled or 59 mm for the 50 mm sled. No original tolerance allowance is silently consumed.
Sleeve clear width 62 / 52 mm, clear height 414 mm, gross axial service package 320 mm. A 3.5 mm face seal land per side produces cap widths 69 / 59 mm and a nominal cap height 421 mm. Use a removable cap with compression bars, not a sliding gate whose full-height parking pocket is omitted from the model. Seal material, land adequacy, clamp packaging and tolerances remain unverified.
Add 20 mm axial package depth ahead of the unchanged 250 mm sled + 50 mm dock: preliminary 8 mm removable-cap/clamp stack, 4 mm seal/seat allowance and 8 mm docking/flush channel allowance. Thus clear inside depth must be at least 320 mm, before any additional tolerance reserve. The service hood is not substituted for this permanent allowance.
Add 4 mm height per rung for top/bottom movement clearance while retaining the original 20 mm guide/partition allocation: rung pitch 434 mm. Four rungs + 200 mm headers + exterior top/bottom walls total 1,976 mm for t=20 or 1,956 mm for t=10, still below 2 m nominal.
Size the detachable sealed hood to at least 110 mm clear width × 450 mm clear height × 350 mm clear projection. It flares outward from the selected slot seal, so the 69/59 × 421 mm removable cap can be parked/manipulated outside the narrow sleeve. It may overhang neighboring CLOSED caps but must not open them. The service case uses a 300 mm working withdrawal stroke, explicitly increasing Rev A's 270 mm: 250 mm body + up to 20 mm front-cassette recess gives 270 mm geometric clearance, then 30 mm additional working travel. This avoids treating cap-closing clearance or full-cap parking as zero volume. Outer hood walls, latches and actual actuator access can require more space. Dock and pre-purge the hood before opening the selected cap; withdraw into the closed hood, replace the selected cell cap from inside the hood, then open the hood to remove the sled. Reverse this with a dry, purged replacement sled and warm-admission sequence. Neighbor sleeves stay shut and do not share an unisolated gas return. Hood structure, glove/manipulation access and support tray are separate service equipment and not yet designed; its volume and mass are outside the unit footprint calculation.
Maintain positive dry-gas flow toward a controlled exhaust. The 110 × 450 × 350 mm hood contributes 17.325 L gross volume; the nominal empty sleeve adds 8.214 /6.889 L, giving about 25.54 /24.21 L for 60/50 mm sled variants, before unmodeled gas lines or glove/actuator pockets. An illustrative ten nominal volume changes at 20 L/min takes about 12.8 /12.1 minutes; ideal mixing would reduce initial contamination to exp(−10), but dead zones, leakage, desorption and gas purity prevent any guarantee from time alone. Require measured local dew point ≤−40°C before −20°C operation, and ≤−60°C before a −40°C qualification excursion, plus validated leakage and purge flow. These are new study targets, not existing qualified system settings. Nitrogen exhaust/oxygen-deficiency controls are required.

The two requested rectangles have only 300 mm inside along withdrawal. Neither fits this 320 mm service package as drawn, even when the lateral count looks adequate. They are short by 20 mm; no fully packaged service capacity, operating power or unit mass is claimed for those combinations. To continue the area-limited study without reducing the inherited sled/dock dimensions, also test 400 × 360 mm, t=20, and 423.5 × 340 mm, t=10. These are explicitly changed envelopes, not the original requested rectangles; both supply 320 mm clear depth.

Table

Variant / outside envelope	Effective pitch	Lateral count / fully packaged slots per rung	Physical / active slots per unit	Unit kW nominal / peak	Unit dry mass Al / Cu	Unit kg/m² Al / Cu
Case 1 + service, 424 × 340, t=20	69 mm	5 / does not fit	Not applicable	Not applicable	Not applicable	Not applicable
Case 2 + service, 450 × 320, t=10	69 mm	5 / does not fit	Not applicable	Not applicable	Not applicable	Not applicable
Case 3 first rectangle + service, 424 × 340, t=20	59 mm	6 / does not fit	Not applicable	Not applicable	Not applicable	Not applicable
Case 3 second rectangle + service, 450 × 320, t=10	59 mm	6 / does not fit	Not applicable	Not applicable	Not applicable	Not applicable
Re-proportioned: 400 × 360, t=20, 60 mm sled	69 mm	4 / 4 conditional	16 / 15	22.82 / 27.32	184.7 / 283.2 kg	1,283 / 1,967
Re-proportioned: 423.5 × 340, t=10, 60 mm sled	69 mm	5 / 5 conditional	20 / 19	28.82 / 34.52	212.3 / 335.5 kg	1,474 / 2,330
Re-proportioned: 400 × 360, t=20, 50 mm sled	59 mm	5 / 5 conditional	20 / 19	28.82 / 34.52	213.5 / 336.7 kg	1,483 / 2,338
Re-proportioned: 423.5 × 340, t=10, 50 mm sled	59 mm	6 / 6 conditional	24 / 23	34.82 / 41.72	241.1 / 388.9 kg	1,674 / 2,701

The re-proportioned cases use bank widths 340 /383.5 mm, nominal areas 0.144 /0.14399 m², and insulation 2.48832 /1.292896 kg. Counts remain conditional on full seal, dock, coupling, cable, stack and structural fit; their 320 mm depth budget has no unallocated positive axial tolerance reserve. The existing 600 mm lane remains a study assumption: a hood with 350 mm CLEAR projection leaves at most 250 mm, less after hood walls/docking structure, outside its projection, which is not a validated operator/cart space. Do not claim this meets service ergonomics or access codes.

Service kit mass is an incremental 1.2 kg per physical slot, whether that slot is active or spare. A 60 mm-sled sleeve estimate is about 1.17 kg: two 414 × 320 × 1 mm side sheets, two 62 × 320 × 1 mm top/bottom sheets, one 62 × 414 × 1 mm rear sheet, one 69 × 421 × 1 mm cap, at 2,700 kg/m³, plus 0.20 kg seals, clamps and purge hardware. Round up to 1.2 kg for both thickness variants; penetration cutouts and actual stiffening are unresolved. This mass allowance is not a pressure-stiffness or seal-force verification.

Rev B addendum — cases 4B and 4C

Sources: expanded Rev B request and later addendum. The later instruction asks for the best Rev B unit power; the earlier case-1-power wording is retained as a sensitivity, not silently substituted. All dimensions here are study allocations, not rated hardware or measured fit. This addendum changes documentation only and does not revise ICD Rev A or authorize PCB work.

Case 4B — shared docking service head on a row rail

Permanent cage hardware is still required. Keep each independently sealed full-depth sleeve, rear sealed utility penetrations, retained front closure/cap, face docking land, purge supply/isolated exhaust, presence/lock confirmation and safe slot isolation. A mobile head is not a substitute for the walls that protect same-rung neighbors. Connect the head and prove seal/purge before opening the selected cap; close and verify that cap before undocking. Units, neighboring compute, backplane power, telemetry and coolant trunks stay connected. Only the selected sled is quiesced and isolated.

For a conservative like-for-like comparison, use the same +4 mm pitch, +20 mm permanent axial depth and +4 mm/rung height as 4A: effective 69/59 mm pitch, required 320 mm inside axial package, and 1.2 kg extra cage hardware per physical slot. Moving the purge/clamp actuators onto the head might later permit a thinner face, but no unsupported reduction is credited. An interface qualification drawing is required before reducing these reservations.

Head and rail allocation: retain the 110 × 450 × 350 mm clear hood volume and choose an 180 mm along-row × 450 mm outward × 550 mm high outer head/tray envelope, plan footprint 0.081 m². The tray supports the sled inside that envelope; do not add another sled-length projection to it. The requested 270 mm minimum stroke is accommodated; retain the more conservative 300 mm working stroke from 4A for the 20 mm front-cassette recess and clearance. Copper-plate sled payload is approximately 12.16 kg before tool loads; guide/coupling extraction forces and the head's own weight remain to be added to actuator/structure sizing.

Use a preliminary 200 mm rail-center gauge measured across the direction of travel, 250 mm carriage length along the row, and 550 mm robot/track band measured outward from the unit face, with a 250 × 550 mm carriage plan, 0.1375 m². Rail gauge is not service-lane width. Add 750 mm clear maintenance/personnel space outside the robot band: 1,300 mm total row service band for planning, with protected/no-person operation in the motion zone. The inherited 600 mm lane does not accommodate this proposed head plus a person. The four 434 mm-pitch rungs need 1,302 mm minimum center-to-center vertical travel; allocate 1,400 mm lift stroke. Rail travel follows the row; allow 0.5 m run-out at each row end. This is not an access-code or guarding approval.

The following is the common per-unit population/mass result for 4B and 4C; their handling systems differ, not their permanent cage interface. Unit mass excludes rail, mast, head, tray and off-unit gas equipment; their loads are not assumed zero.

Table

Source geometry / service configuration	Effective pitch	Lateral / complete slots per rung	Physical / active slots per unit	Unit kW nominal / peak	Unit dry mass Al / Cu	Unit kg/m² Al / Cu
Original Case 1: 424 × 340, t=20, 60 mm sled	69 mm	5 / no axial fit; area also exceeds cap	Not applicable	Not applicable	Not applicable	Not applicable
Original Case 2: 450 × 320, t=10, 60 mm sled	69 mm	5 / no axial fit	Not applicable	Not applicable	Not applicable	Not applicable
Original Case 3: 424 × 340, t=20, 50 mm sled	59 mm	6 / no axial fit; area also exceeds cap	Not applicable	Not applicable	Not applicable	Not applicable
Original Case 3: 450 × 320, t=10, 50 mm sled	59 mm	6 / no axial fit	Not applicable	Not applicable	Not applicable	Not applicable
Area-compliant alternative: 400 × 360, t=20, 60 mm sled	69 mm	4 / 4 conditional	16 / 15	22.82 / 27.32	184.7 / 283.2 kg	1,283 / 1,967
Area-compliant alternative: 423.5 × 340, t=10, 60 mm sled	69 mm	5 / 5 conditional	20 / 19	28.82 / 34.52	212.3 / 335.5 kg	1,474 / 2,330
Area-compliant alternative: 400 × 360, t=20, 50 mm sled	59 mm	5 / 5 conditional	20 / 19	28.82 / 34.52	213.5 / 336.7 kg	1,483 / 2,338
Best service-inclusive alternative: 423.5 × 340, t=10, 50 mm sled	59 mm	6 / 6 conditional	24 / 23	34.82 / 41.72	241.1 / 388.9 kg	1,674 / 2,701

The original rectangles have only 300 mm clear depth: the rail does not fix their missing 20 mm. The 55 mm base-pitch / 59 mm service-pitch variants still require an unverified ≤25 mm complete active stack. No extra throughput or slot-count advantage is assumed merely because the service fixture moves on a rail.

Case 4C — inward-facing ring and rotating central service mast

Main basis: use the best service-inclusive unit, 423.5 × 340 mm, 10 mm walls, 6 slots/rung, 34.82 kW nominal /41.72 kW peak with one spare per unit already included. The highest packing-only Rev B row is 40.82 kW, but its 450 × 320 mm cage does not fit the current service package; show that arithmetic separately as invalid for this service concept. “900 kW” is interpreted as at least 900 kW aggregate nominal unit capacity, not a hard 900 kW input-feed ceiling, accelerator-only power or measured throughput.

Ring geometry: tangential unit width W, radial depth D; each unit's service face is tangent to a regular polygon. Reserve 20 mm minimum shortest inner-corner gap g, and three consecutive empty bays for a personnel entry portal of at least 0.90 m. Occupied unit count is n and polygon bay count m=n+3. Do not model a closed ring with no way for the person to enter the court. The empty entry bays do not contain compute units, so never count them toward 900 kW.

n=\lceil900/P_{\mathrm{unit,nom}}\rceil,\quad m=n+3,\quad\theta=\pi/m

a=\max\left[\frac{W+g/\cos\theta}{2\tan\theta},\ 1.400\right],\quad r_{\mathrm{unit}}=a+D/2,\quad R_{\mathrm{out}}=\sqrt{(a+D)^2+(W/2)^2}

Here a is the inner-face court apothem, not unit-center radius; R_out bounds the actual outer corners. The cosine correction makes g a 20 mm shortest gap rather than only a nominal tangential pitch increment. The free entry width between bounding inner corners for three vacant bays is:

E=2a\sin(4\pi/m)-W\cos(4\pi/m)

Mast/court allowance: mast-column radius 0.20 m, head/tray radial envelope 0.45 m, and 0.75 m radial space for one person servicing stopped/locked machinery give a minimum court radius 1.40 m. The sled stroke is already inside the head/tray envelope and is not added a second time. The requested 270 mm stroke is the minimum reference; the design study keeps 300 mm working travel. Unit-count packing, not this 1.40 m minimum, governs the calculated radii below. This is not a permission for a person to share a moving robot's swept zone.

Rotation alone cannot dock normally to every off-center slot. Require a radial telescoping stage, at least ±0.20 m tangential indexing, and the 1.40 m vertical lift stroke. In the main layout, the nominal maximum dock-point reach from the center is about 2.05 m. Reserve at least 100 mm radial undock/retraction before rotation, subject to a real swept-envelope check: with 0.29 m maximum tangential extent (0.20 m indexing +0.09 m head half-width), approximately 71 mm retraction puts the head's outer corner 50 mm inside the inscribed court boundary. Arm links, cap handling, utility chains and structural deflection can require more; no mast payload/stiffness rating is established.

Area reporting keeps two separate boundaries:

A_{\mathrm{ring+court}}=\pi R_{\mathrm{out}}^2,\quad A_{\mathrm{court}}=\pi a^2,\quad A_{\mathrm{with\ outward\ apron}}=\pi(R_{\mathrm{out}}+1.10)^2

The first is a conservative complete circular allocation including units, central court, entry gap and corner/inter-unit voids; it is not n×unit footprint. The second floor allocation also reserves an outward 1.10 m whole-unit service apron: 0.34 m unit withdrawal +0.75 m person +0.01 m provisional reserve for the service-inclusive cases. This is not a cart-turning, utility-disconnection or legal-egress design. Rare whole-unit removal goes outward, with that unit shut down and its utilities safely isolated/disconnected; ordinary sled service does not require unit movement.

Table

Ring basis / status	Occupied units / total bays	Slots/rung; physical / active per unit	Unit kW nominal / peak	Unit mass Al / Cu; kg/m² Al / Cu	Ring kW nominal / peak	Radii: court a / unit center / outer corner	Floor: ring + court / with outward apron	Entry clear width
Main: best service-inclusive, 423.5 × 340, 50 mm sled; ≤25 mm stack unverified	26 /29	6; 24 /23	34.82 /41.72	241.1 /388.9 kg; 1,674 /2,701	905.32 /1,084.72	2.0395 /2.2095 /2.3889 m	17.93 /38.24 m²	1.328 m
Standard 60 mm sled service alternative; 28.82 kW case-1-power sensitivity, not the original case-1 wall/envelope	32 /35	5; 20 /19	28.82 /34.52	212.3 /335.5 kg; 1,474 /2,330	922.24 /1,104.64	2.4643 /2.6343 /2.8123 m	24.85 /48.08 m²	1.335 m
Highest packing-only Rev B power, 450 × 320; not a service-fit solution	23 /26	7; 28 /27	40.82 /48.92	236.3 /408.8 kg; 1,641 /2,839	938.86 /1,125.16	1.9360 /2.0960 /2.2672 m, arithmetic only	16.15 /35.62 m², arithmetic only	1.401 m

For the main ring, the court's inscribed disk is 13.07 m²; the units' own summed footprints are only 3.744 m². Neither quantity replaces the 17.93 m² ring-plus-court allocation. The actual radial gap between the 0.20 m mast column and the head's inner edge when docked is approximately 1.389 m, exceeding the 0.75 m stopped-maintenance allowance; other joints and fixtures still need collision analysis. Three empty bays give entry tolerance margin; two empty bays would place a 0.90 m portal too close to its nominal minimum for this study.

Power/availability limits: 26 units provide 5.32 kW nominal headroom over 900 kW. With one active sled removed, the instructed arithmetic gives 903.82 kW nominal /1,082.92 kW peak, before service-equipment loads and real standby changes. Removing an entire unit leaves only 870.50 kW nominal; the main ring is not whole-unit N+1 at 900 kW. Twenty-seven such units would be needed to retain ≥900 kW after one whole-unit loss, with a separately recomputed ring. A hard 900 kW plant limit instead requires explicit power capping; the aggregate unrestrained peak is 1.085 MW, not 900 kW. Distributed jobs may still pause/fail during an individual sled removal despite neighbors staying powered.

Unit-only installed mass of the 26-unit ring is approximately 6.27 tonnes with aluminum plates /10.11 tonnes with copper plates, excluding mast, head, rail, services, coolant and external equipment. This is not a structural slab or support-foot load certification. Floor-area and power comparisons exclude external cooling/power infrastructure, nitrogen equipment, head-drive power and building circulation.

Side-by-side comparison — 4A, 4B and 4C

For equal capacity, use 26 of the same conditional best-service units, 905.32 kW nominal, with all permanent hardware and material assumptions unchanged. Rows are one-sided, with 25 inter-unit gaps of 20 mm between 26 units and a 0.5 m run-out/access allowance at each end. Thus row length is 12.511 m: 26 unit widths, 25 gaps and two end allowances. Planning service bands are 1.20 m for 4A (0.45 m fixture +0.75 m person) and 1.30 m for 4B (0.55 m robot/rail band +0.75 m person). Opposed rows, multi-row sharing and multiple-head throughput are not optimized here.

Table

Item	4A — per-slot closure + manually positioned shared hood/tray	4B — same closure + rail-mounted head	4C — same closure + rotating-mast head
Cage hardware cost per slot	+4 mm pitch, +20 mm axial depth, +4 mm/rung height; 1.2 kg per slot	Same; mobile tooling does not replace sealed sleeves or retained caps	Same
Result on original 424×340 /450×320 rectangles	No full axial fit with chosen package; first also exceeds area cap	Same	Same
Best service-inclusive alternative	6/rung; 24 physical /23 active per unit	Same	Same
Unit nominal /peak including320 W	34.82 /41.72 kW	Same	Same
Unit mass Al /Cu; kg/m² Al /Cu	241.1 /388.9 kg; 1,674 /2,701	Same	Same
Head/tray envelope	Shared outer tool allowance 0.18×0.45×0.55 m; removable cap handled in dry hood	Same head; 0.25×0.55 m carriage; 0.20 m rail gauge	Same head; 0.20 m-radius mast +indexing/lift/radial stages
Ordinary sled extraction	300 mm working stroke into hood/tray; unit fixed	Same, head travels along row	Same, pulls INWARD; unit fixed
Person/maintenance space	1.20 m provisional service band	1.30 m provisional band; personnel excluded from motion zone	Court apothem2.0395 m, minimum allowance1.40 m; 1.328 m entry portal; personnel excluded during motion
Equal-capacity floor including normal sled service and ends/court	19.27 m², row allocation	20.52 m², row allocation	17.93 m², ring+court allocation
Equal-capacity floor ALSO reserving outward unit-removal apron	33.03 m²	34.28 m²	38.24 m²
Whole-unit service direction	Rear/outward relative to slot face; controlled shutdown and utility isolation	Rear/outward	OUTWARD, never through the central mast court
Shared failure/queue constraint	Shared hood/tray and operator can serialize service	One head/rail can serialize or block service	One mast/head can serialize or block the whole ring's service

These are allocation comparisons, not an optimum building layout or proof that 4C is denser. Including outward whole-unit access changes the ranking. Machinery foundations, drive/control/gas-service mass and power, rail/mast maintenance clearance, perimeter building aisles and safety/egress requirements remain additional or unresolved.

Additional assumptions and unresolved items — 4B/4C

The shared head uses the same still-unqualified retained cap/sleeve interface as4A. Exact door/transfer-port mechanism, gasket/coupling interfaces, low-temperature purge performance, clamp loads and cap-interlock sequence require drawings and tests; no reduced cage pitch/depth is credited.
Head, carriage and mast sizes are space reservations, not selected mechanisms. Need actuator travel/force ratings including coupling release, 12.16 kg copper sled, head self-weight, moments, utility-chain loads and deflection; validate normal docking across the face. Mast rotation without tangential/vertical axes is insufficient.
The 270 mm user stroke is treated as the minimum sled reference, not an extra length added outside the head. Retain300 mm working travel and include the full cap-handling volume; rail/mast head handling must be verified with a swept model.
Three bays remain empty for a human entry portal. The0.75 m person allowance and0.90 m portal target are study assumptions, not legal egress ratings; determine occupancy, local codes, emergency release, guarding and lockout. The person does not occupy an active mast/rail swept zone.
The1.10 m outward apron and row-end0.50 m allowances are provisional. Confirm full-unit removal cart, lifting route, coolant/power disconnects and cable manifolds; they may enlarge the reported floor allocations.
One installed spare per unit is not an extra spare unit.900 kW means nominal unit input capacity here; close the actual load coincidence,peak duration,service-head supply and whole-unit resilience requirements before facility sizing.
Highest-power units still require the unverified≤25 mm complete stack. The28.82 kW sensitivity provides a standard-thickness alternative; it is not proof that the original424×340,t=20 unit fits the service hardware. Exact25 mm/Gen5/cold qualification and320 W shared-switch adequacy remain open.
Unit mass and kg/m² retain the prior comparison model; ring/rail/service equipment is excluded but must be added for foundation and floor loading. No field-reliability, service-time or uninterrupted-workload claim is made.

Case 5 — aluminum versus copper plate mass

Keep two 410 × 250 × 6 mm plates, 20% pocket/channel void, and the original densities 2,700 kg/m³ Al /8,960 kg/m³ Cu. Dry pair masses are 2.6568 /8.81664 kg. Retaining Rev A's rounded non-copper total gives sled masses 6.000 /12.15984 kg. Permanent service hardware is cage-mounted and does not increase the removed sled mass; it does increase installed-unit mass. No thermal-conductance, pressure-rating, corrosion or thickness-equivalence advantage is assumed from this mass-only material swap.

Table

Area-compliant comparison / configuration	Slots/rung	Physical / active slots per unit	Unit kW nominal / peak	Unit dry mass Al / Cu	Unit kg/m² Al / Cu
423.5 × 340, t=20, 65 mm, no service sleeves	5	20 / 19	28.82 / 34.52	189.5 /312.7 kg	1,316 /2,172
450 × 320, t=10, 65 mm, no service sleeves	6	24 / 23	34.82 / 41.72	212.3 /360.1 kg	1,474 /2,501
423.5 × 340, t=20, 55 mm, no service sleeves	6	24 / 23	34.82 / 41.72	213.5 /361.3 kg	1,483 /2,509
450 × 320, t=10, 55 mm, no service sleeves	7	28 / 27	40.82 / 48.92	236.3 /408.8 kg	1,641 /2,839
400 × 360, t=20, 69 mm, service sleeves	4	16 / 15	22.82 / 27.32	184.7 /283.2 kg	1,283 /1,967
423.5 × 340, t=10, 69 mm, service sleeves	5	20 / 19	28.82 / 34.52	212.3 /335.5 kg	1,474 /2,330
400 × 360, t=20, 59 mm, service sleeves	5	20 / 19	28.82 / 34.52	213.5 /336.7 kg	1,483 /2,338
423.5 × 340, t=10, 59 mm, service sleeves	6	24 / 23	34.82 / 41.72	241.1 /388.9 kg	1,674 /2,701

kg/m² means unit mass divided by its own footprint, not equipment-room structural floor loading. It excludes the service lane, support feet, coolant and external handling equipment. Support-foot reactions and anchorage still require structural design, especially with 12.16 kg copper-plate sleds withdrawing horizontally.

Case 6 — back-to-back mezzanines on ONE carrier

Study only; no module, connector, schematic, PCB or mechanical placement has been selected or changed. This section compares a COM-HPC module on one face of a single vertical carrier with an OAM on its opposite face, each active surface facing its own liquid plate in the sealed sandwich. Earlier side-by-side results remain unchanged. The four-rung formula in the earlier calculation basis applies only to those earlier tables; Case 6 recomputes rung count R. This is not two carriers joined by another connector, and no riser/cable is introduced.

Evidence and governance boundary

The complete governing ICD Rev A was read for this addendum. It does not explicitly prohibit opposite-face modules, but §6.4 requires directly opposed CPU PCIe/accelerator host lane fields, a direct segment, ≤100 mm provisional routed-length goal and hard review at125 mm; correlated channel simulation must demonstrate10–20% loss margin before topology freeze. Opposite-face overlap is not evidence of that alignment. Founder interpretation/approval is required if the final arrangement departs from that geometry; no ICD amendment or architecture approval is implied. §§9.1–9.4 metal load paths, local frames, hard stops and compliant thermal interfaces remain mandatory. Article10 still prohibits placement/routing before approved schematic architecture. Neither this study nor the earlier study resolves the recorded PCB-state conflict.

Official source trail examined for this section:

PICMG ratification notice lists Server Size D160×160 mm and E200×160 mm; Samtec COM-HPC page lists two400-contact connectors and5/10 mm mated heights, ball/column termination options and JSOM/standoff drawings. These are authoritative web findings, but COM-HPC PDF/drawing evidence verification remains unverified: the carrier guide RC2.0d lookup was still indexing after a five-minute wait. Exact carrier plug footprint, alignment-pin penetration, standoff clearance and current production revision remain open; no drawing-based interference clearance is claimed.
OCP OAM r2.0 v1.0 specification confirms a102×170 mm PCB, bottom-side connectors and107 mm connector pitch. Its baseboard component keepout is103×171 mm. These are r2.0 dimensions, not the older165 mm OAM outline.
OAM uses two688-contact Mirror Mezz Pro connectors with BGA attachment. The specification lists218910-1115/218916-1115 and5±0.15/8±0.15 mm stack options. This is not verification of the current schematic J2 or its pin map. The Molex2189101115 page was also examined; supplier PCB drawing extraction was not completed.

Resized envelope, sleeves and integer capacity

Keep the inherited orientation: COM-HPC E's200 mm edge vertical,160 mm along insertion; OAM's171 mm baseboard keepout vertical and103 mm along insertion. The module regions now overlap in projection onto the carrier instead of adding their heights. Retain30 mm combined end margins; removal of the former10 mm inter-module gap is explicit because the modules are on different faces. These margins are study allocations, not vendor keepouts. Thus E-compatible H=max(200,171)+30=230 mm; a D-only H=max(160,171)+30=201 mm sensitivity uses the verified OAM keepout rather than silently treating170 mm PCB outline as171 mm keepout. If only outlines were used, D would be200 mm; that1 mm does not change the seven-rung integer result. Neither shorter carrier proves that connector, stiffener, coolant and converter keepouts actually fit.

Keep 250 mm insertion depth:160 mm module region + unchanged90 mm combined conversion/utility/edge/structure reservation. Do not reduce the90 mm or50 mm rear dock. The shortened carrier has less total utility area despite the same90 mm strip; two DCMs, protection, storage, telemetry and dock escape need a new mechanical/electrical fit study. Rev B permanent sleeve depth remains250+50+20=320 mm, exactly the400×360,t20 unit's320 mm inside depth, with zero positive axial tolerance reserve.

Define hC and hO as COMPLETE carrier-surface-to-active-envelope heights, including mated module spacing, module PCB, populated components/memory, pedestal/TIM allowance and tolerances. The former6 mm rear zone is replaced by a second active stack, not reused as though an accelerator fit in6 mm:

T=3+h_C+h_O+12+4=19+h_C+h_O;\quad p=T+5+4=28+h_C+h_O\quad[\mathrm{mm}]

The3 mm carrier, two6 mm plates,4 mm local structure,5 mm original guide gap and4 mm added sleeve pitch are unchanged allocations. hC=hO=35 mm is a conservative comparison allowance, not a verified OAM or memory height. It gives T89 mm,p98 mm, rather than the old60 mm sled/69 mm pitch. Added underside bolster/standoff structures must fit inside the applicable face allocation or enlarge it; they are not granted an extra uncounted volume.

For outside400×360 mm,t20,B=340 mm; headers total200 mm, exterior top/bottom walls total40 mm. With Rev B20 mm rung structure plus4 mm sleeve-height addition:

q=H+24;\quad R=\lfloor(2000-200-40)/q\rfloor;\quad N=\lfloor340/p\rfloor;\quad K=RN

n_{active}=RN-1;\quad P_{nom}=1.5(RN-1)+0.320;\quad P_{peak}=1.8(RN-1)+0.320\quad[\mathrm{kW}]

Table

Geometry / conditional active stacks	Sled H×depth×T; sleeve pitch	Rung pitch; rungs	Total height incl.headers/walls	Slots/rung; physical / active per unit	Unit nominal / peak
Existing side-by-side,35 mm active +6 mm rear	410×250×60;69 mm	434 mm;4	1,976 mm	4;16 /15	22.82 /27.32 kW
Case6 E-compatible,35+35 mm	230×250×89;98 mm	254 mm;6	1,764 mm	3;18 /17	25.82 /30.92 kW
Case6 D-only,35+35 mm	201×250×89;98 mm	225 mm;7	1,815 mm	3;21 /20	30.32 /36.32 kW
Case6 E,35+25 mm	230×250×79;88 mm	254 mm;6	1,764 mm	3;18 /17	25.82 /30.92 kW
Case6 E,35+20 mm	230×250×74;83 mm	254 mm;6	1,764 mm	4;24 /23	34.82 /41.72 kW
Case6 E,25+25 mm	230×250×69;78 mm	254 mm;6	1,764 mm	4;24 /23	34.82 /41.72 kW
Case6 D-only,25+25 mm	201×250×69;78 mm	225 mm;7	1,815 mm	4;28 /27	40.82 /48.92 kW
Case6 E,45+45 mm adverse stack	230×250×109;118 mm	254 mm;6	1,764 mm	2;12 /11	16.82 /20.12 kW

Main Case6 gains 2 physical slots,12.5%, over the standard side-by-side16-slot unit despite dropping from4 to3 slots/rung; active capacity gains2 sleds/3.0 kW. D-only gains5 physical slots,31.25%, but excludes Size E and is not a substitute for a D/E-compatible envelope. The existing thin side-by-side400×360,t20 sensitivity has20 physical/19 active,28.82/34.52 kW; it still exceeds the conservative E Case6's18 physical slots. The earlier thin10 mm-wall unit has24 physical slots but is a different enclosure/stack basis, not an apples-to-apples Case6 gain.

Integer breakpoints: four positions require p≤85 mm, hence hC+hO≤57 mm; with35 mm CPU this requires hO≤22 mm. Three require p≤113.333 mm, sum≤85.333 mm; beyond that the bank drops to two. Five require p≤68 mm,sum≤40 mm; six require p≤56.667 mm,sum≤28.667 mm. Equality consumes all nominal bank width, not a production tolerance guarantee. A35+25 mm stack at88 mm still misses four positions: thinning10 mm is not automatically another slot. With the E height allocation, six rungs permit H≤269.333 mm, while seven require H≤227.429 mm; the230 mm trial misses seven by2.571 mm. D-only201 mm permits seven, but eight require H≤196 mm. Offset required to clear opposite-face hardware or align lane fields can enlarge H or depth and invalidate these counts.

Service consequences: retain300 mm working extraction; shorter height is not shorter insertion depth. Six254 mm rungs need1,270 mm center travel; seven225 mm rungs need1,350 mm. The existing1,400 mm lift allocation covers either nominal stack if centers are equally pitched, but real offsets/end stops still need checking. For89 mm sleds, the analogous sleeve has91 mm clear width,234/205 mm clear height for E/D,98 mm cap width and241/212 mm cap height. These fit the old110×450 mm CLEAR hood cross-section nominally; no smaller head is credited. The118 mm-pitch adverse sensitivity needs a cap wider than110 mm and therefore a resized hood/head, not the unchanged service fixture. Seal lands/clamps remain unqualified. More rungs require more manifold branches, penetrations and interlocks;320 W shared allowance and fabric port count must be revisited for18/21/24/28 slots.

No Case6 total-mass saving is claimed. Only explicit changed plate geometry earns a credit: E pair230×250×6 mm,20% void gives1.4904 kg Al/4.94592 kg Cu, compared with2.6568/8.81664 kg for410 mm plates; plate-only reductions1.1664/3.87072 kg per sled. All non-plate component masses remain unresolved/unchanged allowances, not proportional to carrier area. If a future unit mass uses the earlier model, stationary hardware must be at least re-tabulated as51+4R=75 kg for six rungs or79 kg for seven, not67 kg; actual resized sleeves, added utilities and reinforcement need a new BOM. More populated slots also add modules/converters. These are not whole-unit mass results.

Conditionally yes in principle, not yet verified for these modules. SMT/BGA connectors on opposite surfaces do not inherently require shared through-hole signal tails. OAM BGA attachment is verified above; Samtec's web-listed ball variants (e.g.ASP-209948-01,10 mm carrier plug) are an unverified drawing lead, not a selected connector. Opposite-face surface pads can share XY projection only if they remain electrically distinct and every escape, via, drill and hardware feature is compatible. Mirroring the module is a three-dimensional orientation change, not permission to mirror a standard's pin numbering or alter its connector pattern.

The dominant checks are not just signal tails:

Bolsters and fastening: OAM's four NPTH mounting holes clear M3.5 screws to a bolster below the system PCB. Its reference screw lengths must be adjusted for baseboard/bolster thickness. On this inverted module that hardware occupies the CPU-facing side; it can collide with CPU connectors, memory, standoffs or thermal structure. The OAM103×171 mm component KOZ drawing is expressly a top-side baseboard accommodation, not a blanket prohibition or blanket permission for all opposite-face electronics.
Alignment penetrations: OAM SMT receiving nuts are soldered at5.7 mm diameter baseboard holes, not zero-volume SMT lands. Map these holes, OAM bolt/tool access and COM-HPC JSOM/standoff screws in a common carrier datum. No coordinates were invented. Exact COM-HPC hardware/tail/locator drawings remain pending. A stagger or local load-spreader window may be necessary; feasibility must include its height/depth penalty.
PCB construction: a through-hole/press-fit alternative, alignment peg or mounting tail may block opposite-face pads or hardware even when the contacts themselves are SMT. Opposing via-in-pad arrays may demand filled/capped vias, blind/buried microvias or staggered escapes; the14–18 layer ICD family is a starting constraint, not a demonstrated dual-sided escape. Require fabricator stack/drill/registration/assembly review. Backdrill access and clearance must be reserved before both-side population; a through-via serving opposite-side pads cannot simply have its functional barrel drilled away. Stub removal, matched structures and possible HDI alternatives need explicit SI/fabricator agreement under ICD§6.3, not a silent waiver.
Assembly/rework: evaluate double-sided reflow retention, connector coplanarity and carrier bow, inspection access and selective rework without remelting/damaging the far connector. Selected module mating order must leave tool access and avoid transferring extraction forces to the opposite module. Module connectors remain internal assembled interfaces, not the live cage service dock.
Thermal press loads: separate metal load paths and hard stops are needed for CPU and OAM on both faces. Balanced nominal forces do not make unsupported PCB or connector housings acceptable load members. Plate flatness, die height, TIM compression, coolant pressure bulge, CTE cycling, electrical clearance and opposed stiffeners must be analyzed together. Written−40°C mated-stack ratings and module Tcase guarantees remain open; no qualification is borrowed from the form-factor standard.

CPU-to-OAM link length and Gen5 risk

Overlap may shorten planar separation but module-outline centers give no host-lane connector position or routed length. Selected pin-map workbooks, connector orientation/escape, lane mapping and mounting hardware decide whether a direct32-differential-pair x16 link fits. Both standard module-to-carrier mating interfaces still exist; a single carrier avoids an additional inter-carrier connector, not those two interfaces. Endpoint module traces, package models and connector loss are not zero.

Illustrative allocation only, not measured geometry or a vendor allowance: reserve60–90 mm total planar carrier escape+route+matching,3–6 mm active vertical via travel,10–18 mm combined mated connector board-separation scale (COM5/10 plus OAM5/8 options; COM drawing verification still pending), and20–60 mm traces on EACH module. Carrier copper would then be 63–96 mm, while the illustrative endpoint-path sum is 113–234 mm, excluding package paths. Connector separation is not the true curved contact length, and physical millimeters in dissimilar structures cannot be summed into an insertion-loss compliance test. These figures illustrate how a carrier route under100 mm can coexist with a much longer full channel; they do not predict this design's channel or prove even the carrier target achievable. Apply ICD100/125 mm review to the agreed routed-length scope and record that scope explicitly, rather than silently treating module traces as free.

Close the design with selected-module/connector S-parameters, package/on-module channels, actual layer/via/backdrill structures, return-path discontinuities and power-plane clearances. Run correlated Gen5 channel analysis for insertion/return loss, crosstalk, skew and margin at the actual operating corners; preserve85Ω stripline and uninterrupted references. No automatic retimer need or retimer-free success is established. An added riser/cable/interposer, a Gen4-only CPU substitution or a geometry waiver conflicts with the present direct Gen5 doctrine unless Founder-approved ICD revision resolves it.

Cases 4B/4C — follow-up system comparison for the FOUR-slot/rung unit

This supplement fills the previously missing system result for the existing 400×360 mm,t20,60 mm side-by-side sled,69 mm sealed pitch,4×4 physical slots. It does NOT replace the earlier best-power thin-sled10 mm-wall tables, and does not use Case6 populations. All results retain one installed compute-off spare per unit and320 W shared/spare allowance. Interpret900 kW as at least900 kW aggregate nominal unit input capacity, not a hard feed cap or application throughput.

Common population, power, mass and availability

n=\lceil900/P_{nom}\rceil;\quad L=nW+(n-1)0.020+2(0.500);\quad b_{out}=D+0.750+0.010

Table

Item, identical for row4B and ring4C	Standard FOUR-slot/rung,t20	Preserved thin SIX-slot/rung,t10
Outside W×D; own footprint	400×360 mm;0.144 m²	423.5×340 mm;0.14399 m²
Rungs; sealed pitch; physical /active per unit	4;69 mm;16 /15	4;59 mm;24 /23
Unit nominal /peak	22.82 /27.32 kW	34.82 /41.72 kW
Required occupied units	40	26
System nominal /peak	912.80 /1,092.80 kW	905.32 /1,084.72 kW
Physical /active /installed spare sleds	640 /600 /40	624 /598 /26
Unit dry mass Al /Cu	184.7 /283.2 kg	241.1 /388.9 kg
Unit own-footprint kg/m² Al /Cu	1,283 /1,967	1,674 /2,701
Installed unit-only dry mass Al /Cu	7.388 /11.330 t	6.268 /10.112 t
Nominal headroom above900 kW	12.80 kW	5.32 kW
After one active sled off, nominal /peak	911.30 /1,091.00 kW	903.82 /1,082.92 kW
After one entire unit off, nominal /peak	889.98 /1,065.48 kW	870.50 /1,043.00 kW
Units required for ≥900 kW after whole-unit loss	41; new row/ring geometry required	27; new row/ring geometry required

Mass calculations use unrounded inherited values184.68832/283.24576 kg and241.092896/388.929056 kg, respectively, then round outputs. No change to the underlying fixed67 kg, insulation2.48832/1.292896 kg, installed sleds or1.2 kg sleeve per physical slot. Aluminum/copper refer only to cold plates. Keep ±30% planning uncertainty. All service machinery, utilities, coolant, foundations and external plant remain excluded, not massless.

Sled-off availability arithmetic retains the instructed shared allowance and subtracts1.5/1.8 kW; actual standby and head-drive loads may differ. Both populations remain above900 nominal after one active-sled outage but neither is whole-unit N+1. The installed spare is not an extra unit, and spare activation overlap needs admission control. Shared rail/mast or purge failure can remove service access or serialize repairs without instantly powering all compute down; no numerical MTBF/availability percentage or uninterrupted distributed-job claim is justified. Peak plant requirements exceed1.08 MW before excluded loads. If900 kW is instead a hard plant ceiling, explicit power capping is required.

4B row floor allocation, reach and travel

Retain a single one-sided row,20 mm gaps,0.5 m at each end,1.30 m service band (0.55 m rail/robot +0.75 m person),0.20 m rail-center gauge,0.25 m carriage length and0.18 m along-row×0.45 m outward×0.55 m high head. Extraction is300 mm INSIDE the head, not another projection. Whole-unit removal is outward/rear relative to the service face, with controlled shutdown/utility disconnection. Apron grows from1.10 to1.12 m when D grows from0.34 to0.36 m.

A_{row}=L(D+1.30);\quad A_{row+apron}=L(D+1.30+b_{out})

Table

Row4B metric	Standard FOUR-slot/rung	Preserved thin SIX-slot/rung
Units; gaps; total row length	40;39;17.780 m	26;25;12.511 m
Unit depth +normal service band	1.660 m	1.640 m
Floor, units +normal service +ends	29.51 m²	20.52 m²
Separate outward whole-unit apron	1.120 m	1.100 m
Floor ALSO reserving outward apron	49.43 m²	34.28 m²
Centered slot extreme across each face	±103.5 mm	±147.5 mm
First-to-last slot-center span along entire row	16.587 m	11.3825 m
Carriage-center range if rails span full L	17.530 m (L−0.250)	12.261 m (L−0.250)
Rung-center vertical span /allocated lift	1.302 /1.400 m	1.302 /1.400 m
Working extraction /head outward envelope	0.300 /0.450 m	0.300 /0.450 m

Slot centers are centered within the340/383.5 mm usable bank; no extra index distance is silently added outside W. The available carriage range exceeds the docking-center span with the stated run-outs. It is a rail space reservation, not proof of stopping distance, actuator reach or cable-chain fit. Lateral carriage positioning provides normal slot alignment; a real head mount must reach the docking face across the0.55 m robot band without treating that band as an additional extraction stroke. Personnel remain outside the moving zone. Lift and head envelope are unchanged because both source sleds remain410 mm high with434 mm rung pitch; Case6's different lift calculation is not imported here.

4C ring floor allocation, reach and retraction

Retain three consecutive empty entry bays,20 mm shortest inner-corner gap,minimum1.40 m court apothem,0.20 m mast radius,0.45 m radial head and0.75 m stopped-maintenance allowance. All variables below are metres. Recompute from the actual W,D,n rather than copying the thin basis:

m=n+3;\quad\theta=\pi/m;\quad a=\max\{[W+0.020/\cos\theta]/[2\tan\theta],1.400\}

r_c=a+D/2;\quad R_o=\sqrt{(a+D)^2+(W/2)^2};\quad E=2a\sin(4\pi/m)-W\cos(4\pi/m)

A_{ring}=\pi R_o^2;\quad A_{court}=\pi a^2;\quad A_{ring+apron}=\pi(R_o+b_{out})^2

Table

Ring4C metric	Standard FOUR-slot/rung	Preserved thin SIX-slot/rung
Occupied units /total bays	40 /43	26 /29
Court apothem a	2.8696 m	2.0395 m
Unit-center radius rc	3.0496 m	2.2095 m
Outer-corner radius Ro	3.2358 m	2.3889 m
Court inscribed disk	25.87 m²	13.07 m²
Sum of unit footprints (NOT ring allocation)	5.760 m²	3.744 m²
Floor ring+court, including empty bays/voids	32.89 m²	17.93 m²
Outward unit-removal apron	1.120 m	1.100 m
Floor ALSO reserving whole-unit apron	59.60 m²	38.24 m²
Three-empty-bay portal E; target≥0.90 m	1.270 m;passes nominal target	1.328 m;passes nominal target
Bay angular indexing	8.3721°	12.4138°
Exact centered slot index required	±0.1035 m	±0.1475 m
Retained tangential stage allocation	at least±0.20 m	at least±0.20 m
Farthest nominal slot dock-point radius	2.8715 m	2.0448 m
Farthest head corner radius while docked	2.8761 m	2.0533 m
Mast surface to head inner edge at face-center dock	2.2196 m	1.3895 m
Vertical rung span /lift allocation	1.302 /1.400 m	1.302 /1.400 m
Working sled extraction	0.300 m inward into head	0.300 m inward into head

For centered slots let x=(N−1)p/2. Dock-point reach is sqrt(a²+x²); docked head outer-corner radius uses tangential extent u=x+0.090 m. For the head corner to remain50 mm inside the court during rotation, nominal face-normal retraction d satisfies:

\sqrt{(a-d)^2+u^2}\le a-0.050;\quad d\ge a-\sqrt{(a-0.050)^2-u^2}

This gives 56.65 mm for the four-slot ring and64.23 mm for the thin ring at their actual extreme slot indices. Retain at least100 mm radial undock before rotation in BOTH. With the full±0.20 m stage allocation rather than just used slot positions,u=0.29 m; required retraction is about65/71 mm for four-slot/thin rings, still below100 mm. Thus the larger four-slot court needs a materially longer arm reach (~~2.88 m head-corner scale versus~~2.05 m), not a larger assumed extraction stroke or a smaller safety gap. The head's inner edge at a centered dock is a−0.45 and its gap to mast surface is a−0.65. Both exceed0.75 m geometrically, but links/utilities can occupy that space. Need normal docking, radial telescoping, tangential indexing and lift; rotation alone is insufficient. A full rotation range and up to180° shortest-path repositioning may be required; no travel time/throughput follows without drive speeds and interlocks.

The retraction inequality only bounds the chosen rectangular head against the inscribed court, not mast/arm links, end effectors, cap parking or cables. Preserve the prior minimum100 mm while requiring actual collision/swept-envelope analysis. Neither portal nor person allowance is access-code certification. Rare whole-unit removal remains outward, never through the mast court, and its apron is counted separately. On the standard four-slot basis the row uses29.51/49.43 m² versus ring32.89/59.60 m² (normal/with apron): ring is not denser under these conventions. On the preserved thin basis it remains20.52/34.28 versus17.93/38.24 m²; including whole-unit access again changes the ranking.

Follow-up verification and open gates

Completed calculation checks: a scripted recomputation checked n=ceil(900/Pnom), RN−1 populations, both stack heights, all Case6 sensitivity rows, row lengths with n−1 gaps, depth-dependent aprons, corner-corrected ring radii, entry width, floor boundaries, exact slot-index reach, retraction inequality, mass totals and single-sled/whole-unit outage arithmetic. Values above are nominal conditional study results, not measured dimensions. Earlier tables and archived Rev A are preserved; no schematic or PCB mutation or review of changed electrical nets was applicable because none were changed.

Warning — possibly wrong until qualified: complete35 mm face stacks, thin25 mm alternatives, opposite-face bolt/bolster clearance, shorter utility real estate, zero axial reserve, seal lands/actuation, mast stiffness, head payload and cold qualification remain unresolved. Selected module CAD/pin maps, connector manufacturing drawings, −40°C evidence, fabricator escape/backdrill review, plate/load FEA, SI correlation and service swept-model tests are specific blockers to implementation, not blockers to the documented arithmetic. COM-HPC PDF evidence remains pending/unverified; no unsupported clearance or Gen5 path claim was promoted to fact. Info: a physically smaller carrier can increase rung count while increasing thickness; it is not automatically a mass or area optimum. Error if implemented without approval: waiving the direct Gen5 geometry, changing the D/E baseline or doing placement/routing before Article10 approval. No such change was made.

New assumptions and unresolved items — Rev B only

Area supersedes square dimensions. The requested 424 × 340 mm case is noncompliant. The trimmed or re-proportioned alternatives are nominal suggestions; certify worst-case outer dimensions, not nominal area alone.
Thin sled remains unproven. Complete CPU/memory and OAM active-side stack must be ≤25 mm. Obtain mated STEP/stack drawings and written −40°C Tcase data; VLP memory alone does not close this. A memory-down server-family lead exists but is nonstandard “D small” and Gen4; no ICD exception or module change is approved.
Neighbors are separated by permanent sleeves, not just doors. Added allowances are +4 mm pitch, +20 mm axial depth and +4 mm per-rung height. Sleeve seals, clamp mechanics, rear penetrations, cold-cycle leakage, pressure limits, differential motion and thermal bridging need detailed validation. The original rectangles cannot fit this chosen service concept with Rev A's 250+50 mm dimensions retained.
Gas service is controlled and measured. A dry external hood is docked and purged before opening the selected sleeve; neighboring gas paths are isolated. The nominal purge-flow/volume and dew-point targets require test validation, exhaust controls and oxygen-deficiency assessment. Cold-plate isolation still needs trapped-coolant expansion/relief provisions during service warming.
Mass basis is retained for comparison. Same aluminum enclosure, 67 kg fixed hardware allowance, memory/module allocations and plate geometry are used; +1.2 kg per slot for the new permanent service kit. Actual larger rung manifolds, switch, copper handling loads, required reinforcements and service-hood mass are not yet known. Shared external hood/tray mass is excluded, not assumed zero. Cooler memory does not earn an invented mass reduction.
320 W shared power is an instructed comparison allowance, not a validated switch budget. Twenty-four- and twenty-eight-slot rows need correspondingly larger fabric port capacity; the earlier 16/20-port example must not be reused unchanged. Switch cooling, bus current, peak duration, fault coordination and cold qualification remain open.
Zero axial reserve is not a production fit. Coupling float, valve/actuator package, full gasket compression, manufacturing tolerances and cable bend radius can require more than the 320 mm service allocation and further reduce slot count under the area cap. The 600 mm lane and hood manipulation/access need a real installation/service drawing; the detachable fixture needs at least 110 × 450 × 350 mm clear volume and the service variant uses 300 mm withdrawal travel, not Rev A's 270 mm.

Rev B verification: the original five baseline case tables, nominal footprint areas, lateral counts, one-spare power, plate/sleeve mass and kg/m² were checked; the addendum extends Case 4 into 4A/4B/4C and adds their side-by-side comparison. Independent review of the addendum confirmed the conditional concept-study scope, ring geometry, head/stroke allocation and per-unit arithmetic. Its bounded row-length finding was corrected to 25 gaps between 26 units, with the four affected floor-area values independently recalculated. Ring calculations explicitly include empty access bays, court area and a separate outward apron, with a shortest-corner gap correction. The service-fixture sizing reserves full-cap handling volume and additional withdrawal travel; unit table mass excludes the external shared equipment. All fit and qualification limits above remain open; review does not approve implementation or amend the ICD.

All unmodified Rev A assumptions, power-path rules, unknown module/fabric interfaces, no-air-cooling doctrine, Class B backplane-family reservation and Article 10 approval gate remain in force. This study is a conditional dimensional/mass comparison, not a claim of live-service qualification.

Archived Rev A baseline

Study revision A — 2026-10-17. Governing document: ICD Rev A, 2026-09-20.

Scope: numerical packaging and architecture study only. No schematic, component placement, board, footprint, or routing change is authorized or performed. This report does not approve the existing schematic or amend the ICD. The previously recorded Article 10 PCB-state conflict is not resolved by this study.

Inputs: study request, governing ICD Rev A, and existing architecture record.

Status convention: “Input” is a user/ICD requirement; “calculated” follows the stated study assumptions; “allowance” is an engineering estimate, not measured hardware; “unverified” needs source/part qualification. Several manufacturer PDF evidence lookups remained pending after waiting. Their provisional values below are explicitly unverified in this session, not absent from their documents. Only the LM5066I facts carrying inline citations have completed datasheet verification here. Consequently this is a conditional feasibility study, not a released mechanical design or qualified BOM.

Coordinates: X = unit width, Y = front-to-back depth, Z = height. Both withdrawal cases are horizontal, with the carrier plane vertical. Front service extracts toward −Y, with slots repeated along X. Side service extracts toward ±X, with slots repeated along Y. No gravity-direction extraction is assumed.

1. Slot pitch and slots per rung

Use the full permitted outside envelope, 380 × 380 mm. Wall thickness t includes the entire insulated wall build, not solid metal. Reserve an additional 10 mm at each end of the slot bank for rails/structure, beyond the walls. Trial sled thickness is 60 mm plus 5 mm slot-to-slot mechanical clearance: 65 mm pitch. These dimensions include no airflow allowance.

B=380-2t-20\ \mathrm{mm},\qquad N=\left\lfloor B/65\right\rfloor

Reserve 50 mm behind the 250 mm-deep sled for stationary recessed male couplings, their float, valves, electrical dock and manifold/cable routing. These are reservations, not sourced coupling dimensions.

Table

Withdrawal case / sensitivity	Wall per side	Clear inside square	Slot-bank width B	Pitch	Slots/rung	Slots/unit	Remaining bank width	Available sled depth after 50 mm dock	Result and assumption
Front, baseline	10 mm	360 mm	340 mm	65 mm	5	20	15 mm	310 mm	Calculated; 250 mm sled leaves 60 mm longitudinal reserve
Front, baseline	20 mm	340 mm	320 mm	65 mm	4	16	60 mm	290 mm	Calculated; leaves 40 mm longitudinal reserve
Front, baseline	40 mm	300 mm	280 mm	65 mm	4	16	20 mm	250 mm	Calculated; zero unallocated longitudinal reserve: marginal, not tolerance-validated
Side, baseline	10 mm	360 mm	340 mm	65 mm	5	20	15 mm	310 mm	Same count after rotating cage/dock 90°
Side, baseline	20 mm	340 mm	320 mm	65 mm	4	16	60 mm	290 mm	Same count after rotating cage/dock 90°
Side, baseline	40 mm	300 mm	280 mm	65 mm	4	16	20 mm	250 mm	Same marginal-depth condition
Thin-package sensitivity, either face	10 / 20 / 40 mm	360 / 340 / 300 mm	340 / 320 / 280 mm	55 mm	6 / 5 / 5	24 / 20 / 20	10 / 45 / 5 mm	310 / 290 / 250 mm	Arithmetic only: requires ≤50 mm-thick sled, NOT demonstrated
Thick-package sensitivity, either face	10 / 20 / 40 mm	360 / 340 / 300 mm	340 / 320 / 280 mm	75 mm	4 / 4 / 3	16 / 16 / 12	40 / 20 / 55 mm	310 / 290 / 250 mm	Arithmetic only: assumes ≤70 mm-thick sled

There is no intrinsic front-versus-side capacity difference in a square unit when the stationary dock rotates with the slots. A fixed manifold/spine, different door build, asymmetric penetrations or lane obstruction can break that symmetry. A 20 mm wall could fit five slots only at ≤64 mm pitch under this model; 65 mm intentionally does not round down to 64 mm.

2. Sled envelope and service motion

Module outlines used for packing are COM-HPC Size D 160 × 160 mm or Size E 200 × 160 mm, and OAM 170 × 102 mm, inherited from the existing architecture study. Exact module drawings, populated-memory height and current OAM mechanical revision remain unverified in this session. Arrange CPU and OAM one above the other in Z, not end-to-end along withdrawal. This is an envelope allocation, not component placement.

Table

Item	Study value	Construction / assumptions	Limit or consequence
Sled outer envelope, H × insertion depth × thickness	410 × 250 × 60 mm	H: 200 mm Size E allocation + 170 mm OAM allocation + 10 mm between modules + 30 mm combined end margins	Covers D/E planar envelopes conditionally; does not establish exact connector orientation or module interchangeability
Depth allocation	250 mm	160 mm module depth + 90 mm combined carrier-edge, conversion, structure and sled-side connector allowance	No overlap proof; female couplings and their sled valves must stay inside this allocation
Thickness allocation	60 mm	3 mm carrier + 35 mm populated active-side stack + 6 mm rear components/clearance + two 6 mm plates + 4 mm combined local sealing/structure allowance	35 mm active stack is unverified, particularly DIMM population; plate support is metallic, never unsupported PCB
Slot pitch	65 mm	60 mm sled + 5 mm guide/assembly clearance	Seal lands, guides and valve actuator must fit without projecting into neighbors
Rung vertical pitch / four-rung stack	430 mm / 1,720 mm	410 mm sled height + 20 mm total guide, partition and structure allowance per rung	Preliminary; partitions must seal independently
Headers and total height	100 mm top + 100 mm bottom; 1,940 / 1,960 / 2,000 mm total	Four 430 mm rungs + two 100 mm headers + two exterior walls of 10 / 20 / 40 mm	40 mm case uses the full 2 m with no extra vertical reserve
Fixed dock reservation	50 mm along insertion axis	Recessed floating male coupling pair, stationary isolation valves, power and signal docking; female pair on sled	Two coolant connections and four isolation functions total, one on each half of each supply/return connection; actual product package unresolved
Travel to clear a flush cage opening	250 mm geometric minimum; 270 mm working extraction stroke	Sled front starts flush at service opening; its deepest moving point is 250 mm inboard; add 20 mm travel allowance	Recessed handles/doors increase travel by their recess; coupling engagement must fit within the defined seated envelope, not be double-counted
Shared lane	≥270 mm swept-depth minimum; 600 mm provisional service lane	600 mm provides withdrawal and some handling space; support extracted sled on a tray/cart	Not an ergonomic, code or accessibility determination; operator/cart geometry may require more
Adjacent-rung service	One sled at a time	Rungs have independent dry-gas volumes, coolant branches and power isolation	Other rungs must remain sealed; no shared removable panel that vents them
Same-rung neighbors	Not solved by a common rung door alone	Opening a shared nitrogen compartment can expose cold neighboring electronics to moisture	Require a sealed service tunnel/airlock, individual sled seals with shutters, or a proven dry-service enclosure; added geometry may invalidate slot count
Bayonet actuation	No whole-sled rotation assumed	Captive rotating collar or cage-side actuator must turn the female bayonet while the sled translates	Actuation access, forces, float and tolerance stack remain unresolved
CPU–OAM host route	ICD target ≤100 mm; hard review at 125 mm	Facing lane fields and direct on-carrier host connection remain required	Envelope fit is not evidence that PCIe Gen5 channel geometry or loss meets ICD §6

3. Rung backplane and scale-up fabric

Answer: conditionally yes as a system, not as a passive shared fabric bus and not for an arbitrary OAM. OAM defines a module interface/envelope; it does not establish that the chosen accelerator implements UALink. Do not equate PCIe, AMD-specific links, Gaudi Ethernet and UALink. A common UALink endpoint/switch protocol and a supported module pin map are required.

Official web research from the UALink Consortium whitepaper reports x1/x2/x4 links, 200 Gbit/s lane class, 212.5 Gbit/s serial PHY and 800 Gbit/s TX plus 800 Gbit/s RX for x4. These numeric specification details remain unverified by the PDF evidence lookup in this session and are used only for provisional lane reservations. Do not interpret 800 Gbit/s per direction as application payload guaranteed to every peer at once.

Table

Fabric question	Numeric reservation / proposed type	Feasibility and limits
Per sled minimum study fabric	1 × x4 full-duplex station: 4 TX + 4 RX differential pairs = 8 pairs = 16 signal conductors, excluding grounds, shield and sidebands	Provisional 800 Gbit/s in each direction. Real accelerators may require many stations; lane count and bandwidth are not frozen
Four-slot rung	4 endpoint ports; 16 full-duplex lanes, 32 differential pairs	Connect every endpoint to a compatible switch; shared switching capacity must be specified
Five-slot rung	5 endpoint ports; 20 full-duplex lanes, 40 differential pairs	Same switched-star principle; not a passive multidrop bus
Sixteen-slot unit	Provision 16 x4 ports, 64 lanes, 128 differential pairs	With one spare per unit: 15 active endpoints, 60 active lanes, 120 active pairs
Twenty-slot unit	Provision 20 x4 ports, 80 lanes, 160 differential pairs	With one spare per unit: 19 active endpoints, 76 active lanes, 152 active pairs
All-to-all reachability / mesh alternative	Switched star is the study baseline; direct full mesh would need N−1 x4 ports per sled	Four-sled mesh: 3 ports/sled, 6 x4 cables; five-sled mesh: 4 ports/sled, 10 x4 cables. Endpoint support for direct links is not established; no full-mesh implementation is assumed
Power connector type	Separate ORv3-style blind-mate positive/return blades, ≥45 A per Class A slot before derating	First-mate chassis and last-mate presence/kill per ICD. Exact MPN, mechanical float and cold evidence open; never carry this power on signal contacts
Future Class B, backplane-only reservation	Preserve connector-family path to approximately 70 A, using additional blades or a double-width slot, per ICD §3.4	No Class B card design, installed hardware or extra Class A power allocation. Exact family modularity, keying and double-width mechanical compatibility remain unresolved; baseline slot count describes Class A only
Fabric connector type	Candidate: 224G-class blind-mate twinax cabled backplane, e.g. Samtec Si-Fly HD SBCM / SBCF family	Manufacturer web material identifies a 224G-class development family; NOT an approved purchasable exact MPN. Mating configuration, package, −40°C rating, availability, mate life and full channel SI remain unverified. No temperature rating is borrowed from a different Si-Fly series
PCB versus cable backplane	48 V busbars + telemetry PCB; separate high-speed cabled dock mounted to the same rung structure	Physical backplane assembly can serve all functions without forcing 200G-class channels through long PCB traces. Standard 64/112G-class connectors are not assumed sufficient for 212.5 Gbit/s signaling
Host PCIe	ICD x16 host link is 32 differential pairs, plus clock/sidebands, on each carrier	Separate from the x4 scale-up reservation; does not cross the rung dock in the baseline
Bridge four rungs	Prefer one unit-wide UALink switch with ≥16/20 x4 endpoint ports and direct cables across all rungs	Four physical rungs need not be four switch tiers. Switch and cable space must fit headers/spine; not demonstrated by 50 mm dock allowance
Cascaded rung switches	No confirmed port count or aggregate bandwidth	Rung-switch-to-unit-switch hierarchy requires explicit switch-vendor routing/management support. Not inferred from a single-tier whitepaper; ordinary Ethernet switches are not substitutes
Inter-rung reach and resilience	Unit height ≈2 m; budget exact cable lengths including bends and endpoints	Do not claim 2 m cable reach from 2 m height alone. Cold cable SI, bend radius, routing and common-switch cooling are unresolved
Hot service	Electrical neighbors may stay powered; shared jobs may still abort	Drain/quiesce fabric before disconnect, isolate the affected port and coolant branch, maintain termination and telemetry integrity. No uninterrupted distributed-job claim

One spare means one per unit, not one per rung. Physical port/power/guide provisioning is never reduced because a slot is temporarily reserved.

4. Unit power with one spare

Treat 1.5/1.8 kW as the total 48 V input envelope per active sled, including its conversion losses. This prevents double-counting DCM losses at unit level. One fully installed cold spare has compute off. Full-load coincidence is assumed; no statistical diversity credit. External refrigeration, pumps, facility power conversion and warm-up heaters are outside this unit-DC figure.

P_{\mathrm{unit}}=(4N-1)P_{\mathrm{sled}}+P_{\mathrm{shared}}+P_{\mathrm{spare,AON}}

Use 320 W illustrative combined allowance, not a specified maximum: 200 W unit fabric switch, 60 W governance/telemetry, 40 W distribution losses, and 20 W spare housekeeping. Actual hardware may exceed this, particularly a high-port-count 200G-class switch; retain the formula until selected.

Table

Case, either withdrawal face	Physical / active sleds	Active-sled nominal / peak	Unit nominal / peak including illustrative 320 W	Unit current: nominal at 48 V / peak at 44 V	Rung power design basis
10 mm walls, 5/rung	20 / 19	28.5 / 34.2 kW	28.82 / 34.52 kW	600.4 / 784.5 A	Fully occupied rung: 7.5 / 9.0 kW; 204.5 A at 44 V peak
20 mm walls, 4/rung	16 / 15	22.5 / 27.0 kW	22.82 / 27.32 kW	475.4 / 620.9 A	Fully occupied rung: 6.0 / 7.2 kW; 163.6 A at 44 V peak
40 mm walls, 4/rung	16 / 15	22.5 / 27.0 kW	22.82 / 27.32 kW	475.4 / 620.9 A	Same electrical basis; tighter mechanical reserve
Rung containing the cold spare	4 active if 5/rung; 3 if 4/rung	6.0 / 7.2 kW; 4.5 / 5.4 kW	Add that rung's actual shared/AON consumption	Raw active-sled 44 V peak: 163.6 A / 122.7 A	Other rungs remain fully populated; do not derate every rung because one unit slot is spare
Full connector-current allocation	All physical slots equipped	45 A × 5 = 225 A/rung, or ×4 = 180 A/rung	45 A ×20 = 900 A/unit, or ×16 = 720 A/unit, plus shared loads	Connector allocation is not an allowed workload power increase	Thermal, return-bus and fault design must consider full physical slot allocation, source limits and interlock policy
All slots accidentally/temporarily enabled	20 / 16 active	30 / 36 kW; 24 / 28.8 kW	Add actual shared loads	Not part of one-spare operating policy	Hardware admission must prevent overlap or plant must carry it; N+1 spare transfer sequence unresolved

The module allocations do not prove that the selected devices will consume 1.5 kW continuously; unit figures are envelope ratings. Peak duration is still undefined. Nearly all sled input ultimately becomes heat, but coolant plant power is additional.

5. Dry mass estimate

First-order allocation, not manufacturer assembly weights. Density assumption: aluminum 2,700 kg/m³; effective insulation 40 kg/m³. All coolant excluded, including liquid inside plates, manifolds and couplings. Baseline includes all 16/20 sleds physically installed, one compute-off spare. No external CDU, service cart, facility cabling, foundations or seismic anchoring.

Plate estimate uses two 410 × 250 × 6 mm aluminum plates with 20% mass removed by internal channels/pockets:

m_{\mathrm{plates}}=2(0.410)(0.250)(0.006)(2700)(0.80)=2.66\ \mathrm{kg}

Table

Assembly / item	Mass allowance	Included basis and sensitivity
Sled cold plates, dry pair	2.66 kg	Calculated from stated slabs/channel void allowance; not a pressure/thermal design
CPU module + memory / OAM	0.80 / 0.60 kg	Unverified assembly mass allowances; exact module weights needed
Carrier PCBA + converters + boot storage	0.70 kg	Allowance including two DCMs and local protection, not just bare laminate
Stiffener, guide shoe, handles and fasteners	0.80 kg	Allowance; load path closes in metal
Sled-side dry couplings/valves/connectors/TIM	0.40 kg	Allowance; cage-side couplings counted with rung, not sled
Per sled total	≈6.0 kg; planning range 4.5–8 kg	Component allowances sum 5.96 kg; rounded to 6.0 kg for population arithmetic
Fully populated 5-slot rung	≈34 kg; rough range 25–46 kg	5 ×6 kg +4 kg stationary rung hardware; range uses 4.5–8 kg/sled and 2.5–6 kg fixed
Fully populated 4-slot rung	≈28 kg; rough range 21–38 kg	4 ×6 kg +4 kg stationary rung hardware; same fixed-hardware range
Unit stationary hardware beyond four rung assemblies	51 kg + insulation	26 kg exterior/interior skins; 10 kg main frame; 6 kg shared dry manifolds; 5 kg spine/switch/cables; 4 kg header fittings. Distinct from rung fixed allowance
Insulation, 10 / 20 / 40 mm wall cases	1.29 / 2.49 / 4.64 kg	Upper-envelope fill volume: [0.38²×2 − (0.38−2t)²×(2−2t)] ×40, with t in metres; small overlap with skins conservatively ignored
Unit, 10 mm walls / 20 installed sleds	≈188 kg dry	4×34 +51+1.29; planning uncertainty ±30% ≈132–245 kg
Unit, 20 mm walls / 16 installed sleds	≈165 kg dry	4×28 +51+2.49; ±30% ≈116–215 kg
Unit, 40 mm walls / 16 installed sleds	≈168 kg dry	4×28 +51+4.64; ±30% ≈117–218 kg
If spare is an empty position instead	Subtract ≈6 kg from unit and its rung	Add back any required blanking/sealing hardware; active power counts unchanged
Material sensitivity	Copper plate pair ≈8.82 kg rather than 2.66 kg	With density assumption 8,960 kg/m³ and identical slab/channel geometry, adds ≈6.16 kg/sled: +98.6 kg/16-sled unit or +123.2 kg/20-sled unit; full copper plates are NOT baseline

Skin allowance approximates a 2 mm outer aluminum skin and 1 mm inner liner. Total 10/20/40 mm walls are insulated assemblies, not solid aluminum. Thermal bridges, vacuum panels, pressure reinforcement or a heavier switch can materially alter mass. The ±30% unit range is a study contingency, not a statistical confidence interval. A 380 mm square, 2 m high body also needs overturning/anchorage analysis, especially during withdrawal.

6. Sled power path and current library parts

ICD §3.7 is controlling: accelerator power bypasses the intermediate 12 V conversion. Current BOM readback contains U18, two U19/U20, and fuse placeholders F14, F15, F18. Their presence is not evidence of final SOA, fault coordination or qualified footprints.

Table

Load / part check	Numeric basis	Conclusion / unresolved limit
OAM high-power input	Protected cage bus 44–59.5 V per ICD, study 900 W nominal /1000 W peak OAM allocation	Takes 48 V directly, behind protection; exact accelerator input/transient limits and cold operation must be confirmed. Never put OAM high-power load through DCM
COM-HPC input	Fixed 12 V, study 300 W nominal /358 W maximum allocation	From regulated local conversion; 25.0 /29.8 A at 12 V. 358 W is ICD ceiling, not measured module consumption
12 V auxiliaries	Study 150 W nominal /300 W peak allocation	Converter-fed; local PoL for lower rails. This provisional reserve is not a newly approved auxiliary load or legacy-I/O expansion
Total converted domain	450 W nominal /658 W peak output, or 37.5 /54.8 A at 12 V	Not 1.5/1.8 kW. CPU+aux allocations and their peak coincidence are assumptions
Always-on and other local input consumption	Study 50 W at 48 V input combined reserve	Isolated AON path must remain independent of payload; this is not a claim that the existing 20 W AON converter can deliver 50 W. Allocate actual AON and other loads separately before release
Converter efficiency assumption	96% for illustrative budget	Converted input 468.75 /685.42 W; loss 18.75 /27.42 W. This is an assumed operating efficiency, not a guarantee at cold/low load
Example total slot input	1,418.75 W nominal /1,735.42 W peak	900/1000 W OAM + converted input +50 W reserve; leaves 81.25 /64.58 W inside the 1.5/1.8 kW envelope. No assertion that a selected module mix actually uses this allocation
DCM3717S60D13K0TN1 power rating	User/library 1,000 W rating; provisional manufacturer-detail lookup: 82 A, input 40–60 V, trim 10–12.5 V	82 A and exact conditions are unverified this session; if applicable, setting 12.0 V gives at most 984 W by current limit, not a blanket 1,000 W. Set regulated output for COM-HPC, not unexamined nominal 12.2 V
One DCM or two?	658 W converted peak vs provisional 984 W at 12 V leaves 326 W unused: 33.1% of capacity, or 49.5% above demand	One is power-capable under this load/thermal assumption; retain TWO installed to comply with the paralleled architecture and existing N+1 intent. Two are not needed merely because total slot input exceeds 1 kW. No different converter is required by these numbers alone
Single-module-failure operation	Surviving converter must support full 658 W /54.8 A output at worst operating condition	Current sharing alone does not prove N+1 availability. Common EN/FLT, input/output shorts, reverse current, ORing and fail-isolation can shut down the array; existing capture is not proven fault-tolerant
DCM overload duration	Web lookup suggests 1100 W, ≤1 ms, ≤10% duty	Unverified detail; no sizing credit used. 1.8 kW slot peak duration is unspecified and cannot be equated to converter pulse capability. Need manufacturer electrical/thermal curves and actual peak waveform
LM5066I basic suitability	10–80 V operating, junction specification −40 to +125°C	Appropriate controller voltage/temperature class; not a rated 45 A power switch by itself. External FETs, shunt, copper, gate stability and timing carry/limit the current
LM5066I current-limit check	CL=VDD threshold 23.4 /26 /28.6 mV min/typ/max ; existing nominal shunt 0.51 mΩ	Calculated limit 45.88 /50.98 /56.08 A, before shunt tolerance/TCR; 1.8 kW at44 V is40.91 A. Minimum threshold has ≈12.2% margin over raw peak; maximum is well above45 A, so protected path must tolerate it until shutdown
Shunt dissipation	0.51 mΩ at40.91 A gives 0.85 W; at56.08 A gives 1.60 W	Kelvin shunt rating, pulse energy, cold drift and thermal path open; nominal ohms alone insufficient
MOSFET SOA / electronic fuse	PWR programming controls external-FET dissipation	Existing provisional FET power limit/timer is not validated SOA. Need input C, source impedance, fault current, cold/hot curves and turn-off timing. No extra monolithic high-current eFuse is assumed: LM5066I + FET stage supplies electronic protection subject to validation
Class-A DC fuse placeholder	No actual fuse MPN or interrupt/time-current/cold rating exists; three instances in BOM	Not adequate for build/release. Replace with specified devices. At least45 A path allocation before derating is NOT a universal45 A fuse prescription. Main fuse and each DCM branch need distinct coordinated ratings
Fuse selection envelope	Continuous slot raw 31.25 A at48 V nominal; peak40.91 A at44 V; main electronic limit up to56.08 A nominal-threshold worst case	Select DC voltage ≥maximum bus including coordinated transients; interrupt rating ≥prospective fault current; I²t and time-current coordination with FETs, branches, connector and copper. Cannot name safe ampere/interrupt rating without plant data
Existing transient conflict	Architecture record flags D1 5.0SMDJ60A versus59.5 V OAM ceiling	Open release blocker carried forward, not repaired or re-verified here. LM5066I's80 V operating range does not protect an OAM against all bus transients
CPU performance conflict found	Official conga-HPC/sILH product material reports PCIe Gen4 and DDR4, not Gen5/DDR5	Unverified datasheet-detail in this session, but a clear vendor-product discrepancy requiring closure. Do not claim this candidate implements the ICD Gen5 host link. Select an industrial Gen5-capable module or obtain explicit Founder-approved ICD revision; no silent downgrade

Manufacturer lookup targets, not substitutes for completed citation verification: Vicor DCM manufacturer product page, DCM PDF, congatec product page, congatec mechanical/user manual, and OAM specification lookup. The stored DCM library reference is Markdown, not a PDF; its existence did not yield a validated PDF citation. The 1 kW DCM is non-isolated per library; it must not be mistaken for the separate isolated AON supply.

7. Unresolved items and input that would resolve them

Table

ID	Item not resolved / affected answer	Numeric provisional value or risk	Required input / closure
U01	Exact CPU SKU, D/E size, memory and populated height	200×160 mm maximum study outline; 35 mm active stack allowance	Orderable −40°C Tcase module SKU, memory SKU, full STEP/keepout and height drawings; temperature warranty
U02	CPU host-link discrepancy	ICD Gen5 versus conga-HPC/sILH web-described Gen4	Manufacturer interface confirmation, alternative Gen5 industrial module, or explicit Founder-approved revision; no automatic change
U03	Accelerator and protocol	OAM does not imply UALink; one x4 station is only a reservation	Exact accelerator SKU/revision, electrical pin map, supported scale-up protocol, station count, host lanes, firmware and cold-operation support
U04	OAM interface in current schematic	Existing J2 is provisional, not a validated OAM r2.0 connector/map	Authoritative OAM pin-map workbook, selected module drawings and exact mated connector MPNs
U05	Bayonet coupling package	50 mm fixed dock; 90 mm depth allowance on sled beyond module	Coupling pair MPN/CAD, recess, mating stroke, radial float, misalignment limits, pressure/flow rating, cold seal material, insertion/rotation force, cycle life
U06	Valve interpretation and emergency isolation	Four isolation functions per sled's two-port interface	Exact integral versus separate valves, manual/automatic operation, closure timing and leak behavior
U07	Common-rung service atmosphere	Cold neighbors cannot simply be exposed to humid room air	Seal/airlock/shutter topology, purge dew point, pressure, nitrogen controls, door-open recovery test and safety assessment; resulting space claim
U08	Shared service lane and gravity support	270 mm working stroke;600 mm provisional lane;≈6 kg sled	Installation drawing for paired units/lane, cart/tray dimensions, operator access and safe handling/anchorage requirements
U09	Wall construction and heat leak	10/20/40 mm total builds, not metal thickness	Panel stack, insulation conductivity, thermal bridges, seals, pressure loads and allowable heat leak;10 mm may not be thermally viable
U10	Envelope/manifold fit	40 mm-wall case has zero spare longitudinal and vertical reserve	Dimensioned structural package, partitions, connector/valve dock, tolerances and cable bends; revise reservations/count if larger
U11	Clamp/TIM/plate design	Two6 mm plates;40–60 mm study boss spacing per ICD	Module load limits, pedestal map, pressure proof design, TIM compression curves, FEA, galvanic/electrical isolation and thermal resistance
U12	Exact mass	≈6 kg sled;≈165–188 kg dry baseline unit	Vendor mass properties and mechanical BOM; skins, rigid frame, coupling and switch mass; shipping and seismic load cases
U13	Fabric connector qualification/availability	Si-Fly HD candidate is a development family, not a selected part	Supplier-released exact mating set, written−40°C rating, 224G full-channel models, mate cycles, float/guidance and procurement confirmation
U14	Switch placement, power and cooling	≥16/20 x4 ports;200 W shared switch allowance not bounded	Actual compatible switch/platform, dimensions, port map, control, power, conduction-cooling path and temperature qualification
U15	Inter-rung topology and bandwidth	Central single switch assumed; no cascade approval	Switch-vendor topology support, application all-to-all bandwidth requirement, contention model, cable lengths/bends and SI validation
U16	Shared workload continuity	Other sleds may stay powered while jobs fail	Fabric quiesce/removal/reset behavior and scheduler recovery policy; acceptable blast radius
U17	Meaning of one spare	Installed cold spare used;20 W standby allowance	Cold/warm/hot or physically empty spare; replacement/admission sequence; whether spare boot overlaps retiring sled
U18	Peak power waveform	1.8 kW duration/repetition not defined	CPU/OAM/fabric simultaneous input waveform, allowed droop, repetition and load-step current
U19	Source and protection coordination	44–59.5 V window; up to56.08 A nominal-shunt worst threshold	Bus plant tolerance/transients, fault-current magnitude/decay, cable impedance, input capacitance, branch coordination and safe isolation policy
U20	Fuse parts	Main and DCM-branch fuse placeholders unqualified	Exact DC fuse MPNs,−40°C evidence, time-current/I²t curves, holders, continuous derating and interrupt ratings against U19
U21	DCM thermal and N+1 validity	One module must sustain658 W output;two installed	Manufacturer validated electrical/derating data at12 V and44 V input, cold plates, load steps, startup/sharing, common-fault behavior and fail-isolation validation
U22	AON and actual per-rail load	50 W reserve is not a20 W-converter rating increase	Per-rail measured/budgeted controller, sensors, memory standby, network and switch loads; verify existing AON capacity
U23	Pending datasheet verification	DCM, congatec, OAM and UALink evidence not fully resolved	Accessible authoritative PDFs/standard revision and completed evidence extraction; these remain unverified rather than undocumented
U24	Governing approval gate	Existing default PCB-state conflict and incomplete architecture approval	Separate approved schematic/ERC remediation and Founder approval; study does not authorize PCB work or amend ICD
U25	Trapped coolant during service and warm-up	Closing supply and return can trap cold liquid in sled/connector cavities	Define coolant chemistry, thermal expansion, trapped volume and service/storage temperature range; calculate worst pressure and provide qualified expansion accommodation or safe relief on every trapped segment. Include system pressure, plate/coupling proof limits, fluid heat capacity/viscosity, design flow, pressure drop and material compatibility
U26	Future Class B backplane-family scalability	Approximately 70 A through extra blades or double-width slot; not a Class A load	Confirm selected connector family's modular blades, contact derating, chassis keying and mechanical space for the future option; no Class B card hardware is authorized

Assumption list

The user-specified outside footprint is taken at its maximum380×380 mm, height≤2,000 mm. Smaller footprints do not inherit these counts.
Four rungs are vertically stacked, each with its own sealed nitrogen compartment; top/bottom headers are100 mm each in this study. Rung partitions/guides consume20 mm per430 mm rung pitch.
Wall t is uniform in X/Y and top/bottom.10/20/40 mm includes skins, insulation and wall build; service door is included in this wall envelope.
Additional slot-bank end structure is10 mm at each bank end. No shared fixed spine consumes extra bank width beyond the stated reservations; if one does, recalculate.
The front and side cases are90° rotations including their docks, guides, manifolds and cable exits. Both withdraw horizontally with a vertical sled plane.
CPU/OAM occupy separate vertical planar regions on one carrier; the200 mm side of Size E is vertical. Exact module connector and thermal keepouts are not represented.
A410×250×60 mm sled is a trial envelope, not a released outline; the5 mm interslot allowance is mechanical, not airflow. No fan, fin, baffle or airflow requirement is introduced.
Both6 mm cold plates are carried by the sled and tied by metallic structure. The “one sealed cold plate over CPU/OAM” requirement is interpreted as one sealed sled cooling assembly compatible with ICD's sandwich plates, using only two external coolant ports. If exactly one physical plate is intended, Founder interpretation is required; no unilateral deletion of ICD sandwich structure.
A35 mm active stack and6 mm rear zone are provisional and must include populated memory, all tolerances and thermal interfaces. A populated DIMM height above this changes pitch and capacity.
All sled-side couplings, valves and handles lie inside the stated seated envelope/flush face. The stationary50 mm dock is sufficient only as an assumption; no exact coupling was chosen.
Male couplings are recessed, floating and cage-fixed; female couplings move with the sled. Supply and return each have isolation on both sides. Rotating collars, not whole-sled rotation, actuate the bayonets.
Minimum full-clear travel is the moving body's inboard depth;20 mm additional travel gives270 mm working stroke.600 mm lane is a starting handling allowance, not validated human access.
A dry-service barrier can eventually preserve same-rung neighbors during extraction without exceeding reservations. This is a critical unproven feasibility assumption, not solved by two coolant shutoffs. If not achievable, live service requirement fails or geometry must change.
PCIe remains local CPU-to-OAM, direct and subject to ICD Gen5/channel doctrine. No Gen4 downgrade, retimer elimination or100 mm path is approved by this packaging study.
Fabric example is homogeneous compatible UALink endpoints, one x4 station per sled and one common unit switch. Arbitrary vendor mixing, passive bus operation, extra station availability and cascaded switches are not assumed.
Fabric pair counts exclude grounds, clocks, management, shields and spare pairs. Each full-duplex lane has one TX pair and one RX pair.
One spare is physically installed per entire unit, compute-off with20 W illustrative AON, not one spare per rung. All slots still carry full physical connector/guide/port provisioning.
The 1.5/1.8 kW values are per-active-sled 48 V input envelopes; all active peaks coincide. No diversity or cold-efficiency credit is used for feed sizing.
The illustrative320 W shared/spare allowance is not a measured or maximum requirement; external pumps, refrigeration, power-shelf loss and chassis warm-up energy are excluded. No facility/AC-input power claim is made.
The 12 V budget uses CPU300/358 W and auxiliary150/300 W; OAM900/1000 W;50 W additional local input reserve;96% conversion efficiency. These are architecture allocations, not selected-module guarantees or new frozen requirements.
One healthy DCM can cover the example12 V load only if current/thermal/voltage/transient conditions permit. Two installed implement parallel/N+1 intent, but fault tolerance is not established by count alone. No peak overload credit is taken.
The0.51 mΩ shunt is nominal; threshold calculations omit resistor tolerance, drift and current-dynamic effects.45 A is a minimum continuous interface sizing requirement, not a maximum possible fault current or a selected fuse rating.
Mass assumes aluminum plates with20% void fraction, specified study densities and all sleds physically installed. Module, coupling, PCB and structural masses are allowances. No coolant or external support equipment is counted.
Unit skin allocation is26 kg, shared frame/manifold/spine/header allocation25 kg, rung stationary hardware4 kg each, and low-density insulation estimated from bounding volume. Additional airlock geometry, copper plates, pressure vessels or a larger switch must be added.
Every unrated component, mated connector, fluid seal, memory device and accelerator still needs explicit cold qualification.−40°C is qualification mode under the ICD, not proof of routine operating permission; admission remains+5°C floor/+7°C release before staged−20°C operation.
Existing schematic/ERC issues and the recorded Article10 conflict remain open. This report changes only project documentation; it neither ratifies an ICD amendment nor permits component placement/routing.
Coolant shutoffs do not by themselves make a removed cold sled safe to warm. Trapped-fluid thermal expansion requires pressure analysis and a qualified expansion/relief provision; no coolant or compliant volume is selected here.
The selected backplane power-connector family must retain the ICD's future approximately 70 A Class B expansion path via additional blades or a double-width slot. Class A population, current sizing and mass estimates do not include a populated Class B option.

Study verification record

All seven question tables and the assumptions were checked against the request and ICD Rev A. Arithmetic checks covered slot counts, rotated withdrawal symmetry, thickness/height stack, extraction stroke, lane/pair counts, one-spare power, dry mass and converted-domain loading. Review corrections added trapped-coolant expansion safety and the future Class B connector-family reservation; source-verification and implementation limits remain explicit. No hardware design mutation was performed.

Rev C Addendum 2 — chip-face routes under R-CL8A / R-CL8B / R-CL8C
Founder rulings — verbatim
Evidence and value-classification register
A. Finished-datum stack budget
B. Gross added chip-path thermal resistance and temperature rise
C. Dielectric planning ratios — not allowable working voltage or hipot
D. Planning recommendation and honest fit classification
E. Coupon matrix and retained provisional gates
Addendum 2 verification and scope receipt
Rev C addendum — Founder ruling R-CL8A: zoned ceramic sandwich plates
A. Finished-datum stack and bounded 4 mm redistribution
B. Thermal penalty: exact ICD basis versus assumed area
C. Bounded process survey and zone applicability
D. Interface, corrosion and manufacturing hazards
E. Qualification gates — provisional engineering targets, NOT ICD-approved requirements
R-CL8A verification and scope status
Rev C — stack evidence, paired rows and facility comparison
1. Complete OAM-face stack height
2. Two facing rows with a shared service lane and whole-unit N+1
3. GB200 NVL72 baseline, limited-boundary ratios and two levels
4. Case6 plate materials and unit dry mass
5. COM-HPC Server D/E CPU survey — information only
Rev C — new assumptions, unresolved items and verification
Rev B — rectangular variants
Calculation basis used in all case tables
Case 1 — 424 × 340 mm, 20 mm walls, 65 mm pitch
Case 2 — 450 × 320 mm, 10 mm walls, 65 mm pitch
Case 3 — both rectangles at 55 mm pitch / 50 mm sled
Case 4A — per-slot sealed service with dry-gas flush
Rev B addendum — cases 4B and 4C
Case 4B — shared docking service head on a row rail
Case 4C — inward-facing ring and rotating central service mast
Side-by-side comparison — 4A, 4B and 4C
Additional assumptions and unresolved items — 4B/4C
Case 5 — aluminum versus copper plate mass
Case 6 — back-to-back mezzanines on ONE carrier
Cases 4B/4C — follow-up system comparison for the FOUR-slot/rung unit
New assumptions and unresolved items — Rev B only
Archived Rev A baseline
1. Slot pitch and slots per rung
2. Sled envelope and service motion
3. Rung backplane and scale-up fabric
4. Unit power with one spare
5. Dry mass estimate
6. Sled power path and current library parts
7. Unresolved items and input that would resolve them
Assumption list
Study verification record