Technical Manual · Cockpit Basis
AI/HPC Data Hall — NVL72 Cockpit — Methodology & Formulas
Every input, equation, constant, output, and reference behind the GB200 NVL72 AI Data Hall Cockpit. All math is anchored to the locked Scenario A basis-of-design and executed by the deterministic engine js/datahall-calculations.js (DATAHALL_CALC) against the deep-frozen model in js/datahall-model.js (DATAHALL_MODEL) — no Math.random() on basis KPIs, no back-solved constants, no per-tab divergence.
01 Purpose & engineering basis
The AI Data Hall Cockpit is a deterministic digital-twin dashboard for a GB200 NVL72 HPC facility operating direct liquid cooling. It answers five engineering questions: what is the IT load per hall and per facility, how is heat split between liquid and air paths, what hydronic flow and how many CDUs are required, what is the honest bottom-up PUE, and what does the electrical distribution look like under the locked load.
Basis of design is locked: the DATAHALL_MODEL object is deep-frozen at page load. No tab can mutate it. Every KPI on every tab reads from the same DATAHALL_CALC.lockedState() and DATAHALL_CALC.pueBasis() call — Rule 1 (one source of truth) of standarization/ACCURACY_VALIDATION.md. The governing hardware reference is the NVIDIA GB200 NVL72 specification; thermal and facility methodology follows ASHRAE A2 and ISO/IEC 30134-2 (PUE) and ISO/IEC 30134-8 (CUEIT).
02 Inputs — LOCKED basis + tab selection
The primary input block is the immutable DATAHALL_MODEL (Scenario A, locked 2026-05-17). Tab-selection controls (scenario toggle, view mode) change which derived quantities are displayed but do not alter the underlying constants. All values below are from BASELINE-DECISION.md and verified in doc-21 worked examples.
| Parameter | Symbol | Value (Scenario A) | Source |
|---|---|---|---|
| Facility data halls | halls | 4 | BASELINE-DECISION.md |
| NVL72 domains per hall | nvl72PerHall | 27 | BASELINE-DECISION.md |
| Physical rack positions per NVL72 | racksPerNVL72 | 2 | doc-21 Ex1 "1 NVL72 = 2 rack concept" |
| Physical rack positions per hall | racksPerHall | 54 | 27 × 2 |
| GPU per NVL72 / CPU per NVL72 | gpuPerNVL72 / cpuPerNVL72 | 72 / 36 | NVIDIA GB200 NVL72 spec |
| IT load per NVL72 pair (LOCKED) | kwPerNVL72 | 132 kW | BASELINE-DECISION.md "Adopted load — 132 kW per NVL72 (NOT per rack)" |
| IT load per physical rack position | kwPerRack | 66 kW | 132 / 2 — doc-21 Ex1 |
| IT load per hall (Scenario A) | itPerHall_kW | 3,564 kW | 27 × 132 — doc-21 Ex1 |
| IT load per facility (Scenario A) | itPerFacility_kW | 14,256 kW | 3,564 × 4 |
| Liquid capture ratio | liquidCaptureRatio | 0.85 | BASELINE-DECISION.md "Liquid capture 85%" |
| TCS supply / return | tcsSupplyC / tcsReturnC | 35 / 45 °C | BASELINE-DECISION.md "35 / 45 °C (ΔT 10 K)" |
| TCS delta-T | tcsDeltaT_K | 10 K | BASELINE-DECISION.md |
| Water density | rhoKgPerL | 1.0 kg/L | doc-21 Ex6 |
| Water specific heat | cpKjPerKgK | 4.186 kJ/kg·K | doc-21 Ex6 |
| LV voltage (line-line) | voltageLL | 400 V | doc-21 Ex3 |
| Power factor | powerFactor | 0.96 | doc-21 Ex3 |
| CDU rating | cdu.ratingKW | 350 kW (EoR) | doc-21 Ex7 |
| Chiller nameplate COP | chiller.nameplateCOP | 6.8 | doc-21 Ex9; Carrier 19DV |
| UPS rating per feed | ups.ratingKW | 4,500 kW (2N) | doc-21 Ex2 |
| Transformer rating | transformer.ratingMVA | 5 MVA | doc-21 Ex4 |
| Active CRAH units | crah.activeUnits | 5 | doc-21 Ex8 |
| CRAH chilled-water ΔT | crah.chwDeltaT_K | 5 K | doc-21 Ex8 |
| PUE design band (target) | pueDesignBand | 1.12 – 1.25 | BASELINE-DECISION.md |
| Data hall geometry (L × W × H) | geometry | 32 × 20 × 4.2 m | doc-21 Ex10 |
Scenario A vs B: Scenario A (LOCKED) treats 132 kW as per NVL72 pair → 3,564 kW/hall. Scenario B (variant, NOT adopted) treats 132 kW as per physical rack → 7,128 kW/hall, which overloads the 4,500 kW 2N UPS (158% loading — impossible per doc-21 Ex2) and requires 12,000 A busway. The cockpit labels every KPI with its basis scenario and locks the derived panel to Scenario A.
03 Calculation methodology
All formulas below are implemented verbatim in js/datahall-calculations.js. Function names are in parentheses. Source lines cite 00-overview-audit.md "Core Calculation Engine" (lines 104-125) and doc-21 worked examples.
IT load — itLoad()
IT_kW = Σ rack_kW + network_kW + storage_kW
In
itLoad(rackPowerKWArray, networkKW, storageKW). For Scenario A: Σ rack_kW = 27 NVL72 × 132 kW = 3,564 kW. Network and storage additive.doc-00 line 107 · doc-21 Ex1
NVL72 power — nvl72Power()
NVL72_kW = racksPerNVL72 × rack_kW
In
nvl72Power(nvl36RackPowerKW, racksPerNVL72). 2 × 66 kW = 132 kW per NVL72 pair.doc-00 line 108 · doc-21 Ex1
Facility load — facilityKW()
Facility_kW = IT_kW + coolingElec_kW + upsLoss_kW + distLoss_kW + aux_kW
In
facilityKW(itKW, coolingElecKW, upsLossKW, distLossKW, auxKW). The denominator for PUE and CUE.doc-00 line 109
PUE — pue()
PUE = Facility_kW / IT_kW
In
pue(facilityKW, itKW). ISO/IEC 30134-2 definition. The result is DERIVED (cyan chip), not the TARGET (1.12–1.25 band from BASELINE). Rule 4: marketing target ≠ derived value.doc-00 line 110 · doc-21 Ex9 · ISO/IEC 30134-2
Liquid heat / air heat — liquidHeat(), airHeat()
Liquid_heat_kW = IT_kW × liquid_capture_ratio (0.85 for NVL72 direct liquid)
Air_heat_kW = IT_kW × (1 − liquid_capture_ratio)
In
liquidHeat(itKW, captureRatio) and airHeat(itKW, captureRatio). Scenario A: 3,564 × 0.85 = 3,029 kW liquid; 3,564 × 0.15 = 535 kW residual air.doc-00 lines 114-115 · doc-21 Ex5
Hydronic TCS flow — hydronicFlowLPM()
Flow_LPM = Q_kW × 60 / (ρ_kg/L × Cp_kJ/kg·K × ΔT_K)
In
hydronicFlowLPM(qKW, rhoKgL, cpKjKgK, deltaTK). Q is the LIQUID CAPTURED heat (85% of IT), not full IT load. The UI must state this basis explicitly per doc-21 Ex6 "UI requirement: show whether flow is based on liquid captured heat or full rack heat."doc-00 line 117 · doc-21 Ex6 · BASELINE-DECISION.md
CDU count — cduRunningCount()
CDU_running = ceil( Liquid_heat_kW / CDU_rating_kW )
In
cduRunningCount(liquidHeatKW, cduRatingKW). CDU_rating = 350 kW (end-of-row, doc-21 Ex7). Recommended installed count is N+1 or N+2 above running minimum.doc-21 Ex7
CRAH residual cooling — fwsFlowM3h()
perCRAH_kW = Air_heat_kW / activeUnits
Flow_m3/h = Q_kW × 3600 / (ρ_kg/m³ × Cp_kJ/kg·K × ΔT_K)
In
fwsFlowM3h(qKW, rhoKgM3, cpKjKgK, deltaTK). ρ = 1000 kg/m³, ΔT = 5 K (CRAH chilled-water loop). Residual air carries the remaining 15% of IT heat. Doc-21 Ex8 demonstrates: Scenario B 214 kW/CRAH → 36.8 m³/h per unit.doc-00 line 118 · doc-21 Ex8
Chiller input — chillerInput()
Chiller_input_kW = cooling_heat_kW / COP_nameplate
In
chillerInput(coolingKW, cop). COP_nameplate = 6.8 (Carrier 19DV, doc-21 Ex9). This COP is NEVER calibrated or back-solved to hit a target PUE. "cooling_heat" in the PUE path equals full IT load (doc-21 Ex9: "cooling heat = 7,128 kW = IT").doc-00 line 119 · doc-21 Ex9
Bottom-up PUE with 5-part basis — pueBasis()
chillerInput_kW = IT_kW / COP (COP = 6.8 nameplate)
pumps_kW = 200 × (IT / 7128) (scale from Ex9 ref)
fans_kW = 250 × (IT / 7128)
cduPumps_kW = 120 × (IT / 7128)
crahFans_kW = 80 × (IT / 7128)
upsDistLoss_kW = 300 × (IT / 7128)
aux_kW = 150 × (IT / 7128)
nonIT_kW = chillerInput + pumps + fans + cduPumps + crahFans + upsDistLoss + aux
Facility_kW = IT_kW + nonIT_kW
PUE = Facility_kW / IT_kW
In
pueBasis(itKW, model). The 6 ancillary reference values (200/250/120/80/300/150 kW) are the Scenario B reference point from doc-21 Ex9, scaled linearly for any IT load. The 5-part basis breakdown (IT / cooling / UPS-loss / dist-loss / aux) is returned for display per doc-00 lines 301-308. Nothing is back-solved.doc-21 Ex9 · doc-00 lines 301-308 · BASELINE-DECISION.md
Electrical — current, kVA, 3-phase kW
I_A = kW × 1000 / (√3 × V_LL × PF) [required current]
kVA = kW / PF [apparent power]
kW = √3 × V_LL × I × PF / 1000 [3-phase real power check]
In
requiredCurrentA(), kva(), threePhaseKW(). Spec uses 1.732 (not full √3 = 1.7320508) to match doc-21 Ex3 exactly. V = 400 V, PF = 0.96.doc-00 lines 122-123 · doc-21 Ex3 · doc-21 Ex4
Room volume — roomVolumeM3()
Volume_m³ = length_m × width_m × height_m
In
roomVolumeM3(lengthM, widthM, heightM). Geometry: 32 × 20 × 4.2 = 2,688 m³ per doc-21 Ex10. Fire-agent sizing (if implemented) requires further inputs per NFPA 2001 — room volume alone is insufficient.doc-21 Ex10
04 Constants & data sources
Every constant lives in the deep-frozen DATAHALL_MODEL with a // source: comment. No economically-material literal lives inside a function body. Key hardware constants:
| Equipment | Model / Reference | Rating / Value | Source |
|---|---|---|---|
| UPS | Modular UPS, 2N A/B topology | 4,500 kW per feed | doc-21 Ex2; BASELINE-DECISION.md |
| Transformer | Cast-resin distribution transformer | 5 MVA (10 MVA conservative) | doc-21 Ex4; BASELINE-DECISION.md |
| Busway | Undersized: 4,000 A → Selected: 6,300 A | ~5,357 A required (Scenario A) | doc-21 Ex3 |
| Generator | Caterpillar 3516E | 2,500–2,750 kW | NVIDIA GB200 NVL72 spec context; BASELINE "NOT 8 MW" |
| CDU | 350 kW end-of-row CDU | 350 kW thermal | doc-21 Ex7 |
| Chiller | Carrier 19DV water-cooled centrifugal | 1,231–4,044 kW; COP 6.8 | doc-21 Ex9; Carrier 19DV nameplate |
| CRAH | Active units per hall | 5 active; ΔT 5 K | doc-21 Ex8 |
| PUE ancillary (Ex9 Scenario B reference) | kW (at IT = 7,128 kW) | Scales by |
|---|---|---|
| Chiller input (COP 6.8 nameplate) | 1,048 | IT / COP (not scaled) |
| Primary pumps | 200 | IT / 7,128 |
| Drycooler / cooling tower fans | 250 | IT / 7,128 |
| CDU pumps | 120 | IT / 7,128 |
| CRAH fans | 80 | IT / 7,128 |
| UPS + distribution loss | 300 | IT / 7,128 |
| Auxiliary (lighting, controls, security) | 150 | IT / 7,128 |
| Total non-IT (Scenario B reference) | 2,148 |
05 Outputs
| Output | Engine function | Unit | Basis chip | Interpretation |
|---|---|---|---|---|
| IT load per hall / facility | lockedState().itPerHall_kW | kW | BOD LOCKED | 27 × 132 kW (Scenario A). Never randomised. |
| Liquid heat | lockedState().liquidHeat_kW | kW | DERIVED | IT × 0.85; drives CDU count and TCS flow. |
| Residual air heat | lockedState().airHeat_kW | kW | DERIVED | IT × 0.15; drives CRAH sizing. |
| TCS total flow | lockedState().tcsFlowTotal_LPM | LPM | DERIVED | Based on liquid captured heat (85%), not full rack heat — label required. |
| TCS flow per rack position | lockedState().tcsFlowPerRack_LPM | LPM | DERIVED | Total / 54 rack positions. Shows 80.4 LPM for Scenario A. |
| CDU running count | lockedState().cduRunningCount | count | DERIVED | ceil(liquid_kW / 350). Recommended installed = N+1 or N+2 above running. |
| CRAH flow per unit | lockedState().crahFlowPerUnit_m3h | m³/h | DERIVED | Per active CRAH at 5 K CHW delta-T. |
| Required busway current | lockedState().requiredCurrent_A | A | DERIVED | Must not exceed busway rating. 5,357 A Scenario A → 6,300 A selected. |
| Transformer apparent power | lockedState().transformer_kVA | kVA | DERIVED | IT_kW / PF. 3,713 kVA → 74.3% of 5 MVA (Scenario A, plausible). |
| UPS loading % | lockedState().upsLoadingPct | % | DERIVED | IT_kW / UPS_rated. 79.2% Scenario A — plausible with margin. |
| PUE (bottom-up) | pueBasis().pue | — | DERIVED | Facility / IT. Shows the honest derived value, never the target. |
| PUE 5-part basis | pueBasis().basis | kW | DERIVED | IT / cooling / UPS-loss / dist-loss / aux — human-auditable breakdown. |
| Data hall volume | dataHallVolume() | m³ | BOD LOCKED | 32 × 20 × 4.2 = 2,688 m³ per hall. |
06 Worked example — Scenario A (LOCKED basis)
Numbers reproduce the engine output exactly for Scenario A (132 kW / NVL72). Examples 1–10 are the acceptance tests in tools/test-datahall-calc.mjs (57/57 assertions). Below is a compressed trace of the ten examples.
- Ex1 — IT load & rack count: 27 NVL72 × 2 racks = 54 rack positions/hall. 27 × 132 kW = IT = 3,564 kW/hall; 66 kW/rack position.
- Ex2 — UPS sizing: IT 3,564 kW ÷ UPS 4,500 kW = 79.2% load — plausible (2N, some margin). Scenario B 7,128 ÷ 4,500 = 158.4% — impossible.
- Ex3 — Busway current: I = 3,564 × 1000 / (1.732 × 400 × 0.96) = 5,357 A → 4,000 A busway undersized; 6,300 A selected.
- Ex4 — Transformer loading: kVA = 3,564 / 0.96 = 3,713 kVA; 5 MVA load = 74.3% — plausible.
- Ex5 — Cooling split: Liquid = 3,564 × 0.85 = 3,029 kW; Air = 3,564 × 0.15 = 535 kW.
- Ex6 — TCS flow (Scenario A): Flow = 3,029 × 60 / (1.0 × 4.186 × 10) = 4,342 LPM total; 4,342 / 54 = 80.4 LPM/rack. Flow basis = liquid captured heat (85%), not full rack heat.
- Ex7 — CDU count: ceil(3,029 / 350) = ceil(8.65) = 9 CDUs running; recommended installed 10–12.
- Ex8 — CRAH residual: 535 kW / 5 CRAH = 107 kW/unit; m³/h = 107 × 3600 / (1000 × 4.186 × 5) = 18.4 m³/h/CRAH.
- Ex9 — Chiller input & PUE (Scenario B reference point): Chiller = 7,128 / 6.8 = 1,048 kW. Non-IT = 1,048 + 200 + 250 + 120 + 80 + 300 + 150 = 2,148 kW. PUE = (7,128 + 2,148) / 7,128 = 1.30. For Scenario A (IT 3,564 kW scaled): PUE ~1.30 by same methodology.
- Ex10 — Data hall volume: 32 × 20 × 4.2 = 2,688 m³/hall. Fire-agent design requires additional NFPA 2001 inputs; room volume alone is insufficient.
PUE reading: The derived PUE (~1.30 at full load, Scenario B reference) reflects chiller-based direct liquid cooling without free-cooling bypass. The target band 1.12–1.25 (BASELINE-DECISION.md) is achievable with partial free cooling or heat rejection improvements but is a design target, not the operational result from this engine. Rule 4 (ACCURACY_VALIDATION.md): target ≠ derived value — the cockpit shows them in different colours and chips.
07 Accuracy validation — 6 rules
Codified in standarization/ACCURACY_VALIDATION.md and enforced by tools/probe-accuracy-validation.mjs (40/40 acceptance tests). Gate any change that touches the cockpit pages.
| Rule | Requirement | Failure example |
|---|---|---|
Rule 1 — One source of truth | Every KPI on every tab reads from DATAHALL_CALC.lockedState() or pueBasis() — never from a separate variable or inline literal. | Dashboard showing 66 kW/rack while the electrical tab shows 132 kW/rack simultaneously. |
Rule 2 — No Math.random on basis KPIs | IT load, PUE, Facility load, GPU count, CDU count, UPS/transformer ratings: deterministic. Reload 20× — must be identical every time. | Random jitter on TCS flow or dashboard rack-load causing different values each load. |
Rule 3 — Every metric carries a denominator | PUE = Facility ÷ IT; UPS loading = load ÷ rated; transformer = kVA_load ÷ kVA_rating. Denominator shown in tooltip or inline. | PUE shown without stating the IT kW basis. |
Rule 4 — Target ≠ derived value | Target PUE ≤ 1.12 (design band, blue chip) and derived PUE 1.30 (engine output, cyan chip) are two different numbers with different colour meanings. | Showing "PUE 1.12" (target) in green as if it were the operating result. |
Rule 5 — Terminology matches engineering basis | Use "NVL72 domain" (132 kW logical pair) and "physical rack position" (66 kW per slot). Never call a 66 kW rack an "NVL72 rack" — NVIDIA spec says ~120 kW per NVL72 rack-scale system. | Labelling the 66 kW slot as an "NVL72 rack" when it is one of two physical positions in an NVL72 footprint. |
Rule 6 — Basis chip on every critical KPI | Each KPI carries: MEASURED / DERIVED / TARGET / SIMULATED / BOD LOCKED / MANUAL. Click opens a drawer with formula, inputs, output, scope, denominator, source. | A PUE tile with no chip — operator cannot determine if the value is a target or a derived result. |
08 References & standards
- NVIDIA GB200 NVL72 — Blackwell GPU rack-scale specification: 72 GPU + 36 Grace CPU per NVL72, direct liquid cooling, ~120–140 kW per NVL72 domain (adopted design value 132 kW).
- BASELINE-DECISION.md (locked 2026-05-17) — Single immutable source of truth for the whole cockpit: 4 halls, 27 NVL72/hall, 132 kW/NVL72 (Scenario A), 85% liquid capture, TCS 35/45 °C, PUE design band 1.12–1.25.
- 21-calculation-worked-examples.md (doc-21) — 10 fully-traced examples (Ex1–Ex10) that are the acceptance test cases for
tools/test-datahall-calc.mjs(57/57 assertions). - 00-overview-audit.md (doc-00) — Core Calculation Engine definition (lines 104-125), Source Sanity References, equipment corrections (Cat 3516E ≈ 2.5–2.75 MW).
- ASHRAE A2 / ASHRAE 90.4 — Inlet temperature envelope for HPC cooling; energy efficiency for data centres.
- ISO/IEC 30134-2 — Data centre key performance indicators: Power Usage Effectiveness (PUE) — definition, measurement scope, and boundary.
- ISO/IEC 30134-8 — Carbon Usage Effectiveness relative to IT equipment energy (CUEIT).
- Carrier 19DV centrifugal chiller — Nameplate COP 6.8 used in doc-21 Ex9 (not calibrated to hit a target PUE).
- Engine sources —
js/datahall-calculations.js(DATAHALL_CALC, pure derivation engine) +js/datahall-model.js(DATAHALL_MODEL, deep-frozen basis). - standarization/ACCURACY_VALIDATION.md — 6 accuracy rules + 40 acceptance tests enforced by
tools/probe-accuracy-validation.mjs.