BESS is customer-supplied. Optional for firm baseload (nuclear). EMS supervises but does not specify. Profiles: Tesla Megapack (AC), CATL EnerOne (DC).
Supermicro SYS-421GE-NBRT-LCC, 4U, 8× B200 SXM, ~10.5 kW per node sustained. MGX/GB200 deferred.
ARCNODE ships Compute Container + Grid Container. Thermal Module retired — CDU rack-mounted inside Compute. Grid Container required for DC-coupled, conditional for AC-coupled (obsolete when Tesla Megablock ships).
All containers 10ft high-cube ISO (interior 2,680mm). SafeGear (2,413mm) does not fit standard. Uniform fleet.
Trihal secondary 415Y/240V (not 480V). 60A 3-phase rack PDU = 43 kW per unit. 4× PDUs per Compute Container = true 2N at 80 kW. HGX PSUs accept 240V phase-to-neutral.
hardware_selector_map.yaml in edp-api retired. Profile→asset selection moves into manifest.yaml as a profiles: section, authored in this repo where variant decisions originate. edp-api consumes the manifest as the single hardware contract — flat URL maps + curated profile mapping in one fetch. Prevents drift between adding a new variant and exposing it to a profile.
Each container assembly has its own local origin (floor center, X = container length toward mating face, Y across, Z up). Cross-container coupling is explicit: each container module exports named cq.Location constants for its mating frames (e.g. CG_MATING_FRAME on compute, with mirrored normal on grid). A higher-level deployment assembly composes containers by aligning frames. No shared global origin between containers.
Each plate variant has cad/specs/{plate_id}/spec.yaml mirroring the equipment-spec pattern: schema_version, deployment_contexts (commercial / defense_forward), penetration_schedule, outer_dims, mating_pair, provenance. Single source for plate metadata feeding model build, BOM generator, drawing generator, and penetration-schedule artifact. Manifest exposes it as plates.{id}.spec_url.
Two contract artifacts pinned: per-assembly bom.yaml (split sections — parts: equipment list with qty, plates: with id+version) expresses one container's BOM and is multiplied by container count downstream; bom.json is a flat line-items array with a procurement_path discriminator (catalog | custom_fabrication) and a top-level metadata block (deployment_id, profile, manifest_version, generated_at). No hierarchy in either.
Step 4 builds compute-container/commercial-ac only. Grid container is treated as a notional mating frame (named coordinate constant per ADR-007), not a built assembly. Grid container assembly lands in step 6.1, which is the first end-to-end validation of the mating-frame contract. commercial-dc compute-container assumed identical to commercial-ac pending step 6.8 verification — if they diverge there, that's a real second variant; if not, commercial-dc becomes a pointer.
v1 ships single mutable URL s3://arcnode-artifacts/manifest.yaml with an in-file version: field populated by semantic-release. No per-version URL pattern, no MANIFEST_VERSION config in edp-api, no startup version check. Revisit when the schema stabilizes (post step 6) and dev/staging environments split. Risk being accepted: an in-flight edp-api job reading the manifest mid-update sees a torn read; mitigation is per-job fetch + in-memory pin for job duration.
ARCNODE-default Trihal sub-config (GRD-XFM-001):
- 13.8 kV class (utility-typical US MV; matches GRD-SWG-001 SafeGear)
- Off-load taps ±2.5%/±5% (standard, ~50 mm top adder)
- Side-entry HV + LV terminal boxes (saves height vs top-entry)
- Louvered IP31 ventilated top
That config lands at 1,580 × 820 × 1,860 mm. Container interior height 2,680 mm (ADR-004) gives 820 mm overhead, which exactly meets Schneider's natural-convection guideline. The 4× louver cutouts on long walls (Q5) augment but are not load-bearing for that guideline.
Procurement-time check: confirm the procured Schneider SKU encodes this sub-config. Alternate sub-configs (top-entry terminals add ~140 mm height; on-load tap changers add more) push past the natural-convection budget — those would require forced ventilation in addition to louvers, captured as a separate ADR if a different sub-config is procured.
All v1 plates (CG, BG-AC, BG-DC, CD) use radially-slotted bolt holes at the 4 corners + 2 long-axis midpoints (6 of 8 perimeter bolts). 13 mm slot × 11 mm Ø. Short-axis midpoints stay round.
Per-bolt thermal offset analysis (6061-T6 plate vs A36 receiver frame, Δα = 11.9e-6/K, commercial ΔT = 85 K):
| Bolt position | Distance from pattern center | Radial offset @ 85 K | Margin (round Ø11, 0.5 mm radial clearance) | After ISO 2768-m fab tol |
|---|---|---|---|---|
| Corners (±260, ±360) | 444 mm (half-diagonal) | 0.449 mm | 0.051 mm | NEGATIVE |
| Long-axis midpoints (0, ±360) | 360 mm | 0.364 mm | 0.136 mm | 0.036 mm |
| Short-axis midpoints (±260, 0) | 260 mm | 0.263 mm | 0.237 mm | 0.137 mm |
Round corner holes fail outright; long-axis midpoints have only 36 µm net margin (positive but too tight to ship). Short-axis midpoints have ample margin and can stay round.
Two mitigations were considered:
- Option 1 — slot the affected holes (adopted): the constraint is bolt-to-slot-end clearance ≥ worst-case displacement. Min slot length = D_bolt + 2·(δ_thermal + δ_fab + δ_margin) = 10 + 2·(0.449 + 0.1 + 0.2) = 11.5 mm; rounded up to 13 mm for round-number fab stock + extra safety. Per-side travel (slot_length − bolt_diameter)/2 = 1.5 mm covers both corners (0.449 mm budget = 0.749 mm with fab tol + margin) and long-axis midpoints (0.364 mm budget = 0.664 mm). Same slot length serves both. ~$5–10/plate fab cost delta. Commercial headroom 0.75 mm; defense headroom 0.61 mm even at MIL-STD-810H ΔT = 111 K (-40 to +71 °C).
- Option 2 — tighten to ISO 2768-f: drops fab tolerance from ±100 µm to ±50 µm. Buys back 50 µm everywhere; 15–30% fab-cost premium per plate forever. Brittle under wider operating ranges.
Option 1 is structurally robust (eliminates the constraint, not just shrinks fab uncertainty) and future-proof (defense ΔT works without rework).
Slot length is encoded in mounting_bolts.slot_length_mm in each plate's cad/specs/{plate_id}/spec.yaml. Geometry logic in cad/model/_plate.py::_cut_mounting_bolts (corners + long-axis midpoints get slots, short-axis midpoints get round holes). Sim asserts the slot accommodates δ_thermal + δ_fab + δ_margin per side (sim/cg/test_run.py::test_slot_accommodates_thermal_offset). Derivation lives in theory.ipynb "Design risk mitigation" cell.