================================================================================ INTERFACE CONTROL DOCUMENT -- COLD-NATIVE CARD-CAGE COMPUTE PLATFORM ICD Rev A -- September 20, 2026 Program: Cold-Native Vertical Server Carrier (design engine project of record) Authority: Founder-approved Architecture and Cold-Power Design Study Rev B (Sep 20, 2026). This ICD is the standing constitution for every board designed in this program. Every design session inherits it. Changes require explicit Founder approval and a revision letter. NOTE: This document is written vendor-neutral and platform-neutral by intent. ================================================================================ 1. PROGRAM CHARTER (summary) 1.1 The platform is a sealed, dry-nitrogen, sub-zero, conduction-cooled vertical card-cage compute system. Boards are clamped between liquid cold plates (sandwich architecture). There is no air cooling anywhere. 1.2 Program board set, in build order: (1) power spine / bus interface, (2) card-cage backplane, (3) supervisory governance controller [first fabricated prototype], (4) Slot Class A compute carrier, (5) supporting citizens as defined. 1.3 Performance doctrine: maximum sustained compute throughput and minimum host-to-accelerator latency. The supervisory controller may limit performance for safety; the carrier power tree shall not be the normal performance limiter. 1.4 Design-for-validation doctrine: simulation retires design risk; hardware retires qualification risk. No performance or reliability claim is made beyond the validation rung actually completed. 2. ENVIRONMENT 2.1 Operating ambient (enclosure): -20 C. Study excursion: -40 C (qualification mode only until every part and process is qualified). 2.2 Atmosphere: sealed dry nitrogen. Condensation control (dew point, purge, door-open recovery, moisture sensing) is a chassis requirement. 2.3 Cooling: conductive only. Cold plates contact module heat spreaders, die pedestals, and converter thermal lands. The PCB is never the primary heat spreader. 2.4 Orientation: vertical, blind-mate insertion into card cage. 3. BUS AND POWER DEFINITION (RATIFIED) 3.1 Cage bus: 48 V DC nominal, OCP/ORv3-style blind-mate busbar interface. 3.2 Module compatibility window: 44-59.5 V (OAM r2.0 high-power input); plant nominal and transients coordinated below the 59.5/60 V ceiling and hot-swap/TVS limits. 3.3 Slot Class A (frozen): 1.5 kW nominal, 1.8 kW peak. Card interface sized for at least 45 A continuous at 44 V low-line (raw 1.8 kW low-line current 40.9 A) before connector derating. 3.4 Slot Class B (backplane placeholder only): ~3 kW future slot. Reserve connector-family path to approximately 70 A via additional power blades or double-width slot. No Class B card hardware is designed, installed, or sized on Class A cards. 3.5 Connector doctrine: separate positive/return power blades; protective chassis contact mates FIRST / breaks LAST; presence/power-kill contact mates LAST / breaks FIRST; independent management contacts. 3.6 Protection stack per card: slot/source fuse coordination, hot-swap controller (80 V class, external FETs, PMBus telemetry), UV/OV, inrush limiting, reverse-energy blocking as required, hardware kill independent of host firmware. 3.7 Conversion doctrine: 48 V delivered DIRECT to the accelerator module. Regulated 12 V generated locally only for the CPU module (fixed 12 V input standard) and auxiliaries, via paralleled conduction-cooled converter modules. Accelerator power is never converted to 12 V. 3.8 Always-on domain: isolated housekeeping rails (12 V AON, 5 V, 3.3 V) powering the supervisory controller independent of payload state. 4. COMPUTE MODULE BASELINES (RATIFIED) 4.1 CPU: COM-HPC Server Type Size D/E module, extended-temperature SKU explicitly ordered with -40 C minimum Tcase rating (industrial Xeon D class per Rev B citation). Commercial 0 C-minimum modules are a later variant behind qualification. 358 W module input ceiling. 4.2 Accelerator: OAM r2.0, vendor-agnostic. Carry AMD Instinct-class (750 W, 54 V) and Intel Gaudi-class (up to 900 W) envelopes until module selection. NVIDIA is out of carrier scope except as a purchased vendor baseboard, which enters the cage as its own citizen. 4.3 Vendor power stages remain on their modules. The carrier never recreates module-internal voltage regulation. 5. WARM ADMISSION AND THERMAL GOVERNANCE (RATIFIED) 5.1 Admission floor: all boot-critical sensors at or above +5 C. Release threshold +7 C with 2 C hysteresis. 5.2 Sequence: always-on domain only at insertion; enclosure validation (dry-gas indication, no leak/condensation alarm, coolant flow, plausible sensors, plates present); warm-soak to threshold; enable protected 48 V payload; enable conversion and standby rails; release CPU reset; verify telemetry; enumerate accelerator; then authorized staged cooldown to -20 C in bounded steps under rate-of-change, minimum-temperature, rail-margin, and link-error-counter supervision. 5.3 Any sensor below validated floor, telemetry loss, flow loss, leak indication, or rail excursion: throttle, controlled shutdown, payload isolation. Hardware kill path functions without host firmware. 5.4 Warm-up energy (heaters or warm coolant) is a chassis function; the card commands and verifies, never supplies. 5.5 The supervisory controller must start reliably at -40 C even though compute admission waits for +5 C. 6. HIGH-SPEED AND STACKUP STANDARD (RATIFIED, PRE-LAYOUT) 6.1 Laminate: Megtron-6-class ultra-low-loss family with HVLP copper (preferred R-5775(N)/R-5670(N); halogen-free R-5375(N) alternate if fabricator-qualified). Low-temperature coupons mandatory. 6.2 Stackup family: 14-18 copper layers; at least four solid ground references; dedicated 48 V and 12 V distribution layers; paired stripline for Gen5 escape. Final count follows module pin maps and fabricator field solve. 6.3 Channel doctrine: PCIe Gen5 x16 host link, 85 ohm differential stripline, no plane splits, back-drilled vias, matched via structures. 6.4 Retimer-elimination geometry: CPU-module PCIe edge placed directly opposite the accelerator host-link connector, lane fields facing inward; one direct segment, no riser, no cable, no intermediate connector. Provisional routed-length goal 100 mm or less, hard design review at 125 mm, subject to mechanical feasibility; validated by correlated channel simulation with 10-20 percent loss margin before the topology is frozen. 6.5 Passenger doctrine: single boot NVMe only; no general storage expansion; no legacy I/O; no peripheral controllers that backplane management can serve. 7. TELEMETRY MAP (MINIMUM SENSOR SET) Each converter bank; each module power-entry region; CPU pedestal edge; accelerator pedestal edge; coolant inlet; coolant outlet; both plate edge regions; card leading edge; backplane connector region; one board-center reference. Distributed digital sensors plus module telemetry plus redundant analog trips. Sensor accuracy below 0 C is a budgeted item (IC accuracy, gradients, attachment, calibration). 8. COLD-RATED BOM LAW 8.1 No liquid aluminum electrolytic capacitors anywhere. Conductive polymer, polymer tantalum, film, and MLCC arrays; every MPN verified to category minimum of -40 C or below. 8.2 MLCC: X7R/X8R or C0G/NP0; capacitance verified across DC bias and -40 C; minimum 2x nominal design allowance. 8.3 Timing: TCXO or qualified XO rated -40 C or below; no bare crystals. 8.4 RTC: backplane-powered always-on preferred; else -40 C supercapacitor or relocated energy source. No coin cells in the sealed enclosure. 8.5 Connectors: every mated stack verified by written -40 C rating (plastics, plating, contact force, cycle life). 8.6 Conformal coat: low-modulus silicone class with verified thermal shock; connectors and thermal interfaces masked. 8.7 Inductors, shunts, precision networks: cold behavior verified per part (permeability, TCR drift at thresholds, mechanical bonding). 8.8 Assembly: lead-free alloy and land patterns validated by thermal-cycle coupons; the dominant reliability risk is repeated CTE strain, not steady cold. 9. MECHANICAL DOCTRINE 9.1 Clamp force closes through metal spacers/standoffs and perimeter through-bolts, never through unsupported PCB. 9.2 Perimeter stiffener frame plus local load-spreader frames at module connectors; no pedestal load through mezzanine connectors. 9.3 Study boss pattern: perimeter 40-60 mm spacing, reinforced at module long edges and symmetric about each die pedestal; final pattern by module drawings and plate FEA. 9.4 Die pedestals carry compliant replaceable thermal interface with hard compression stop; parallelism, flatness, and connector stack analyzed together. 9.5 Vertical extraction loads (guides, handles, blind-mate forces, cold-line couplings, plate mass) are chassis design inputs. 10. GOVERNANCE AND GATES 10.1 Layout gate: no component placement or routing until the schematic architecture is approved against this ICD. 10.2 Sourcing rule: temperature ratings and standards claims carry part numbers or citations; anything unverified is logged as an open assumption, never asserted. 10.3 The Rev B assumptions register (45 items) is the standing validation job-jar; retiring items follows the validation ladder (simulation, component cold-soak, eval boards, stackup coupons, governance controller prototype, carrier). 10.4 Air-cooling conventions are deleted design rules. Any reappearance is a defect. 10.5 This ICD is amended only by Founder-approved revision. Sessions cite the ICD revision they were built against. ================================================================================ ICD Rev A end. Program constitution effective September 20, 2026. ================================================================================