Approach and contribution
30-second case brief
From rack heat to cooling continuity.
- Problem
- Coordinate heat paths, airflow, capacity redundancy and cooling power through normal operation and failures.
- Contribution
- Developed the heat-balance, component-power, route-availability and coupled thermal-storage models, with inspectable interactive tools.
- Result
- Approximately 3.0°C lower inlet with containment in Case A; uninterrupted modeled coolant flow through the 60-second backed transfer in Case B.
ApproachSI heat/flow balances · component availability · event-aware integration · energy conservation · native EnergyPlus reconciliation.
Project scope and responsibility
- Context
- Interactive engineering training · cooling-system balances and four-zone air mixing
- Role
- Model formulation, engineering scenario design, calculation workflow, and interactive technical presentation
- Configuration
- Eight cooling architectures · generic and Vera Rubin NVL72 rack scenarios
- Design basis
- 24 × 8 kW generic racks or 330 kW NVL72 · 110% cooling sizing allowance
- Methods
- Heat and flow balances · component power · thermal storage · event simulation
- Evaluation
- N / N+1 / N+2 / 2N / 2(N+1) · maintenance and failure states · four-zone airflow model
Cooling lab: 192 kW generic operating heat with a 110% sizing allowance, or the NVL72 preset. The integrated network comparison uses that configuration. Power-transfer and airflow exercises retain their named input cases.
Cooling systems and load basis
This interactive training studio follows rack heat through air or liquid cooling, equipment sizing and failure response. Start with the cooling architecture and operating load, then compare five engineering arrangements under one shared scenario. Equipment cards expose duty and standby groups, isolation points, power and control dependencies. Planned maintenance, unexpected failure and failure during maintenance are separate test conditions. The adjacent assessment relates the selected test to Tier I–IV objectives without assigning a Tier classification. Design-point flows, electrical demand and temperature checks remain expandable below the comparison. Eight architectures remain available, including a Vera Rubin NVL72 hardware preset based on NVIDIA’s published reference.
The generic starting case is 24 × 8 kW = 192 kW of operating IT heat. A separate 110% design allowance gives 211.2 kW of net IT-side cooling sizing duty; whole modules are rounded up from that requirement. This chosen allowance is a training input, not a universal code requirement. The operating-load control can represent part load or overload without silently changing the installed design. One NVL72 rack uses a 330 kW cabinet design basis. A selected 98% liquid capture divides it into 323.4 kW liquid and 6.6 kW residual air, with a 495 L/min published design-flow floor.
Outdoor dry bulb and dry-cooler approach determine whether the requested water supply can be produced. Water-to-air and interloop approaches then check whether that supply can cool the selected load. Required flow remains a hypothetical sizing value when a temperature path is infeasible; achieved electrical demand and heat rejection are then unavailable. Standby equipment restores service only if its physical temperature path and distribution remain available.
Interactive workspace · Cooling systems, engineering arrangements and failure response
Interactive training studio · calculated scenarios
Follow the heat.
Test the resilience.
Choose a cooling architecture, inspect its heat and flow balances, then test the cooling paths under a failure or maintenance event.
Choose a rack basis, then inspect the cooling architecture and response sequence.
NVIDIA reference specifications · checked 28 October 2026
Vera Rubin NVL72
The preset uses NVIDIA's DSX cabinet TDP as a design basis. Load fraction, 98% liquid heat capture, water-reference properties, 500 kW CDU/plant blocks, 50 kW air units, power factor and response timings are editable training inputs. Shelf ratings describe equipment capacity separately from scenario heat and operating power.
NVL72 uses direct-to-chip cooling. Select the generic scenario to explore the other cooling architectures.
CRAH · Chilled-water room cooling
Room air crosses a chilled-water coil. Fans return cooled air to the racks; the facility-water loop carries heat to the plant.
24 racks. Generic modules represent complete net cooling paths. The NVL72 case sizes the warm-liquid plant, CDU and independent room-air modules separately. Enter the net IT heat that one module can remove at the selected operating point, after its auxiliary heat is accounted for. The design allowance defaults to 110% and stays at or above 100%; N rounds this sizing duty up to whole modules. Operating load changes the heat currently generated, independently of the design allowance.
Outdoor temperature sets the dry-cooler feasibility boundary. The entered chiller/DX COP is a fixed training input, not a weather-dependent equipment curve. Try 30°C and 40°C outdoors with the same warm-water setpoint.
One scenario · five engineering arrangements
Compare what stays connected.
Capacity reserves, independent distribution and uninterrupted cooling answer different questions. Set one event below; every arrangement uses the architecture, loads, weather and module duties selected above.
Connections, isolation and operating provisions
These are explicit scenario assumptions, not properties guaranteed by a redundancy label. A checked provision means it is included in this network model. On one-route arrangements it cannot create a second route.
In a dual-route test, maintenance removes the selected item on B; the injected fault acts on A. In a single route, both act on that route. A module under maintenance consumes a spare first. Planned maintenance prepositions available replacements; an unexpected fault waits for the pickup delay. A shared header or common cooling-power loss disables both routes until the event ends. No generator or UPS is assumed here. A common power event keeps cooling off for the source interruption plus the entered restart delay.
Equipment and connections
● Connected◇ Ready spare× Isolated / unavailable
Performance on the same time axis
| Arrangement | Minimum usable capacity kW IT | Minimum actual removal kW IT | Shortfall kWh | Peak air bulk rise K | Cooling duty restored elapsed s | Air bulk recovered elapsed s | IT curtailed kWh |
|---|
Flow interruption: Not evaluated. Module availability is not a hydraulic or airflow-continuity calculation.
Liquid / chip temperature: Not evaluated. Use the separate coupled-loop power-transfer mode for its stated boundary.
Outdoor heat rejection during faults: Not evaluated. The design-point balance below is separate from this transient dispatch model.
Equations, boundaries and interpretation
N is the whole number of duty modules required for the 110% design-sizing duty (or the edited allowance). N+1 and N+2 add one or two spares to every nonempty group on a single route. 2N duplicates the required groups on A and B. 2(N+1) gives each route its own spare in every group. A shared spare 2N+1 arrangement is not modeled.
Usable capacity is the maximum total IT heat supportable at the selected air/liquid split, after serial bottlenecks and connection states. Actual removal is limited by active IT heat in each circuit. Cooling shortfall = active IT heat − actual removal. Its integral is reported in kWh; it is not automatically stored heat. Rack shutdown reduces active demand and is reported separately as curtailed IT energy.
For the air circuit only, stored heat follows E(t + Δt) = max[0, E(t) + (air shortfall − available air recovery duty) × Δt]. Air bulk rise = E / effective air-side thermal capacity. This effective well-mixed store is not a rack-inlet or chip temperature prediction. Spare connected capacity can remove this store after pickup. Liquid shortfall is recorded as unremoved heat without assigning it to a fictitious flowing liquid inventory. A 100% liquid case therefore has no air-temperature result.
Net module duties are assumed available at the selected setpoints. Outdoor, air-coil, rear-door and CDU temperature approaches gate the relevant circuit. A feasible temperature check does not replace equipment performance curves or site design. The separate design-point balance includes auxiliary heat; the route transient tracks net IT heat only. The models share edited inputs, not unverified transient power or outdoor-rejection results.
All delays, spare dispatch, isolation behavior, loads and effective thermal capacities are training assumptions. The default 192 kW case uses three 80 kW duty paths and a 30 s pickup delay. Adding a spare can shorten a deficit without guaranteeing continuity. No cooling-capacity label alone establishes Tier III or IV.
resilience-workspace-20261030-v1 · Equipment counts, connection assumptions and interval heat balances are included in the export.
Design-point heat paths, flows, electrical demand and assumptions
- 01Server air
- 02CRAH coil
- 03Facility water
- 04Outdoor heat rejection
Air side
- Required server airflow
- —
- Required residual room-cooling airflow
- —
- Required external cooling-fan power
- —
Liquid and plant
- Required facility-water flow
- —
- Required technology-liquid circulation
- —
- Evaporation rate (two-phase)
- —
- Required pump / compressor power
- —
Rack power, coolant and cooling groups
| Cooling group | Design heat · kW | Capacity per module · kW | N required | Installed modules |
|---|
Inspect and edit the engineering assumptions
The 110% sizing allowance, 80 kW generic modules, 500/50 kW liquid/air modules, 30 s response, 150 kPa pressure rise, 65% efficiency, COP 4 and 3 kW outdoor auxiliary are selected demonstration values. They are not measured equipment performance or claims about common industry values; the NVL72 reference card identifies the sourced manufacturer values. Air uses ρ = 1.2 kg/m³ and cp = 1.006 kJ/(kg·K); facility water uses ρ = 1,000 kg/m³ and cp = 4.18 kJ/(kg·K). Enter the actual fluid properties for a selected liquid. The single-phase immersion preset uses illustrative dielectric properties of 1.6 kJ/(kg·K) and 800 kg/m³.
Flow requirements include the pump heat entering each liquid circuit. External fan heat enters the cooled air; pump work enters its respective fluid. The circuit rejection includes compressor work. Outdoor auxiliary electricity dissipates outside the fluid circuits and is added separately to the total environmental heat release. If a temperature approach is infeasible, required flows remain hypothetical sizing requirements and achieved operation is unavailable. In the NVL72 case the warm facility-water flow carries liquid-captured heat, while a separate low-temperature room-air unit uses the entered COP. Server-integral fans are already inside IT power. Rear doors are passive in this calculation; their server airflow carries the full rack heat. The two-phase value is an evaporation mass rate, not a pumping flow.
Compare the eight architecture requirements
One boundary · eight architectures
Compare the heat paths.
All rows use the current IT load, pressure rises, efficiencies and plant settings. Single-phase immersion uses the dielectric preset; two-phase immersion uses the entered latent heat. The table shows design-point requirements.
| Architecture | Air / liquid heat kW | Server air m³/s | Facility water kg/s | Technology flow kg/s | Electrical demand kW | Plant condition |
|---|
cooling-studio-20261030-v4 · Design-point balances with a linked five-arrangement network comparison. Inspect assumptions for the complete model boundary.
NVL72 cooling through a power transfer
Case B holds IT heat at 330 kW through an assumed independent IT supply while the cooling plant loses power. Generator availability after 45 seconds and another 15 seconds of restart create a 60-second bridge. The continuity case begins at 40°C TCS supply, rather than the architecture preset’s 45°C steady design point, to examine thermal headroom below the published inlet reference.
Backing only the technology pump preserves flow but concentrates heat in the 250 L technology inventory. Backing both loops connects that inventory to a 2,000 L facility buffer, and the separate air unit removes the remaining 6.6 kW. The calculation tracks the two heat stores and finite UPS energy independently, so a large buffer cannot conceal a stopped pump or depleted battery.
Interactive tool · NVL72 backup power and coupled-loop thermal storage
Interactive NVL72 power-transfer example
Keep the liquid moving.
Keep its temperature controlled.
Compare a cooling interruption, a backed circulation pump, and backed circulation with thermal storage and residual-air cooling. Follow the coolant during the generator transfer.
The NVIDIA DSX reference ↗ describes GPU cooling using loop thermal mass during power outages. Uptime Institute's Tier explanation ↗ connects fault tolerance to the complete infrastructure. This calculation examines continuous cooling using stated power and thermal assumptions.
Cabinet heat uses the 330 kW DSX design basis per rack. The example holds IT power constant through a separate assumed IT supply. Liquid capture (98%), water-reference properties, inventory, backup capacity, pressure loss and response times are selected training inputs. Technology flow uses at least 1.5 L/min per cabinet design kW; changing operating heat alone does not reduce this design-flow floor. The continuity case starts at 40°C TCS supply, providing 5 K below the published inlet reference. The architecture preset above uses a separate 45°C steady design point. Each tool names its initial conditions explicitly.
Assess flow, flowing supply temperature and residual-air service separately. A lower displayed temperature during an interrupted-flow case does not indicate better cooling: no supply or chip temperature is predicted during the gap. The shaded interval is the plant power outage, not necessarily a coolant-flow outage.
Fluid, power and thermal-storage inputs
Battery energy available to the cooling circuit = nominal kWh × usable fraction × conversion efficiency. Flow requires an energized pump circuit; thermal volume alone does not maintain circulation. With both loops circulating, heat is exchanged between the technology inventory and facility buffer through the modeled CDU. With only the technology pump running, that heat accumulates in the smaller technology-loop inventory.
The two well-mixed inventories obey CT dθT/dt = Qliquid + PT − QCDU and CF dθF/dt = PF + QCDU − Qplant. C = volume × density × specific heat. Temperature deviations θ are relative to the initial operating state. QCDU = Qliquid + PT + G(θT − θF) while both loops circulate; G is the entered per-rack effective coupling conductance multiplied by rack count. The TCS supply rises only with its own stored energy. Initial FWS temperature is derived from the same conductance and running duty.
The 65 kW/(K·rack) CDU conductance, 250 L and 2,000 L inventories, 170 kW recovery margin, 45 s generation delay and 15 s restart are selected training values, not equipment specifications or representative industry defaults. The plant rejects no liquid heat during transfer. After restart, it removes running duty plus a temperature-controlled recovery load, capped by the entered margin. Facility-volume sizing numerically solves this coupled model for the 45°C TCS reference, over the displayed simulation window, assuming adequate UPS energy and zero transfer gap. The window covers at least five minutes and at least three minutes after plant restart. A blank sizing result means no solution within the entered model's 20,000 L search range or no temperature headroom. Bulk-loop temperatures describe this mixed-inventory calculation; gaps indicate no circulation. The residual-air unit has an independent condenser auxiliary.
Same rack load. Three cooling-power arrangements.
| Power arrangement | No TCS flow · s | Peak flowing supply · °C | Above 45°C · s | Liquid / air unserved · kWh | Nominal UPS required · kWh |
|---|
Air distribution and rack-inlet temperature
Case A resolves the distribution of heat and supply air across four analytical rack zones. The server airflow is sized at the 192 kW design heat and retained as the operating load changes; this is different from the operating-flow sizing shown in the architecture tool. Supply-air allocation and recirculation determine local inlet mixing, while a cooling-capacity shortfall stores heat in the hall.
In the separately named 85%-load airflow exercise at 20°C supply, the chosen 25% open-aisle and 4% contained recirculation inputs reduce the hottest modeled inlet from 23.40°C to 20.43°C. A duty-unit outage increases the contained case’s transient peak to 23.66°C before recovery. At the reference fan setting, proportional server heat and server airflow make all four inlet temperatures equal. The additional 40%-fan example exposes unequal supply allocation: the least-supplied zones recirculate more hot exhaust. These comparisons use identical rack loads, fan settings and a 30°C outdoor condition, showing sensitivity to the selected analytical mixing assumptions. These are calculated comparisons, not measured or CFD results. The main architecture exercise starts at 100% operating load with a separate 110% sizing allowance.
Interactive tool · Airflow, containment and rack-inlet temperatures
Interactive tool · calculated thermal scenarios
Operate the data hall.
Change the load, supply temperature, airflow, and containment. Compare rack-zone inlet temperatures, cooling margin, and a four-minute unit outage.
Hottest inlet · 15-minute event
How the calculation works +
Four rack zones use sensible heat balance and selected recirculation fractions (open 25%, partial 12%, contained 4%). The resulting approximately 3 K containment difference is an analytical mixing sensitivity to these inputs, not measured or CFD evidence. Equal zone inlet values occur when the proportional server flow and heat give the same rack temperature rise and every zone receives enough air. Lower fan speed exposes the unequal air-delivery allocation. The mixing rise is r/(1-r) × rack temperature rise. Capacity and COP follow explicitly assumed outdoor-temperature curves. The transient combines a 40-second mixing response with a 6 MJ/K effective hall heat capacity. Zone colors show calculated zone averages. The 80 kW modules, outdoor derate of 1.2% per K above 25°C, COP curve, 40 s mixing response and 6 MJ/K thermal store are selected training values, not common industry or manufacturer performance.
The temperature comparison uses the ASHRAE A1-A4 recommended dry-bulb inlet range of 18-27°C.
Equations, inputs, scenario table, and referencesSeven reference cases and model equations
Six reference scenarios
| Scenario | IT load | Unit-state capacity margin | Peak inlet (15 min) |
|---|---|---|---|
| Contained · normal | 163.2 kW | 62.4 kW | 20.43°C |
| Open aisles · normal | 163.2 kW | 62.4 kW | 23.40°C |
| Contained · unit outage | 163.2 kW | -12.8 kW | 23.66°C |
| Contained · standby replacement | 163.2 kW | 62.4 kW | 20.43°C |
| Contained · unequal delivery at 40% fan | 163.2 kW | 62.4 kW | 27.61°C |
| Contained · full design load | 192.0 kW | 33.6 kW | 20.50°C |
| Contained · 120% load / unit outage | 230.4 kW | -80.0 kW | 27.89°C |
Calculated from the shared model and reference inputs. Capacity margin refers to the selected unit state; peak includes the complete event.
ΔT rack = Q IT / (ρ × cp × server airflow)
T inlet = T supply + r / (1 - r) × ΔT rack
r = r baseline + (1 - r baseline) × max(0, 1 - supply flow / server flow)
Q IT = ρ × cp × fresh airflow × (T exhaust - T supply)
dT bulk / dt = (heat load - heat removed) / C hall
P fan = active units × 5 kW × speed³
Simulation and BIM evidence
A native EnergyPlus 24.2 benchmark checks the fixed 163.2 kW airflow-reference load in a 400 m², 3 m-high adiabatic zone. A fixed 163.2 kW internal gain operates for a 24-hour design day, and an ideal sensible-cooling system holds the zone at 20°C. The native hourly output reconciles against 3,916.8 kWh of supplied IT heat. This check isolates the heat balance; fan, compressor and annual weather effects are separate model boundaries. The software handoff below also regenerates equipment and load schedules from the edited labs and supplies an unrun air-side EnergyPlus input. The completed fixed benchmark stays distinct from those editable inputs.
Simulation and BIM workflow
Connect calculations to system design.
Geometry, plant operation and thermal response require different engineering models. The comparison below explains the inputs and outputs associated with each software role. The downloadable schedules provide a common numerical basis with explicit units and assumptions.
Each export regenerates equipment duties and counts, separate air/liquid hourly inputs, weather and water setpoints, and the selected event traces. Revit uses the equipment schedule; IES VE and eQUEST use the load and condition schedules for manual configuration. An editable EnergyPlus IDF contains an ideal air-side sensible balance for the selected air load and outdoor design point. These files are inputs, not completed native runs. Liquid heat remains in its separate schedule. The power-transfer and four-zone exercises retain their own explicitly named cases.
Software roles, common inputs and design outputs
Revit
Coordinate the physical system
- Model inputs
- Rack and cooling-equipment schedules, electrical demand, pipe connections, duty/standby groups and access zones.
- Engineering output to produce
- A coordinated BIM layout, equipment schedules and distribution diagrams.
IES VE
Evaluate annual plant operation
- Model inputs
- IT-load profiles, separate liquid and room-air loads, water temperatures, plant efficiencies, weather and control sequences.
- Engineering output to produce
- Hourly plant energy, heat rejection and cooling performance under the selected weather file.
eQUEST
Benchmark conventional air cooling
- Model inputs
- The generic air-cooled cases, zone sensible loads, schedules, fan power and DX or chilled-water efficiencies.
- Engineering output to produce
- A conventional air-system annual-energy benchmark for comparison with the same load basis.
EnergyPlus
Build a reproducible system benchmark
- Model inputs
- The common heat loads, schedules, plant assumptions and separate air/liquid heat balance, with explicitly modeled components.
- Engineering output to produce
- A native input model and hourly results reconciled against the reference heat balances.
The interactive tools use the linked analytical models. The native EnergyPlus output above is a separate sensible-load reconciliation. Revit, IES VE and eQUEST describe complementary software roles; the fixed one-rack schedule package supplies their common input basis. Its schedules contain equipment loads, flow requirements and event traces for comparison across BIM and energy models. A CFD analysis uses geometry, airflow boundaries and heat loads to resolve spatial temperatures and hotspots.
Design criteria and standards
Equipment documentation sets the selected rack’s operating reference; environmental guidance sets the air-inlet comparison. Infrastructure standards organize different assessments, including energy, safety, maintenance and fault tolerance. The applied register below ties each criterion to a calculation, input or required design record. Published editions identify the reference documents; adopted codes and permit requirements depend on the actual site jurisdiction.
Design basis · standards and evidence
Connect each requirement to a design decision.
The review starts with the project jurisdiction and selected equipment. Each reference below identifies the evidence needed for its assessment. The current lab supplies heat balances, flow requirements and capacity-event comparisons.
| Check | Criterion / source | Evidence | Interpretation |
|---|---|---|---|
| Case A rack inlet | 27°C upper recommended dry-bulb reference (ASHRAE A1–A4) | 20.43°C contained normal; 23.66°C contained outage peak | Below the selected dry-bulb reference. Humidity and equipment class are separate environmental checks. |
| Case B TCS design flow | NVIDIA DSX ≥1.5 L/min per design kW | 330 kW × 1.5 = 495 L/min | Design-flow floor applied; fluid properties and pressure loss remain explicit inputs. |
| Case B inlet condition | NVIDIA published 45°C liquid-cooling design point | 40°C initial TCS; 41.85°C peak in 5 minutes | Below the selected inlet reference in the backed case. This reference is not an automatic plant-trip condition. |
| Cooling power continuity | Maintain flow, temperature reference and residual-air cooling through transfer | 0 s no-flow; 0 kWh air shortfall; 0.195 kWh nominal UPS for a 60 s bridge | All three selected cooling functions maintained in the reference calculation. |
| Infrastructure topology | Uptime Tier objectives and, separately, ANSI/TIA-942-C | Matched-delay component/route tests, including planned isolation and a shared dependency | Cooling outcomes are evaluated by function. Whole-site Tier certification and TIA rating follow their own assessment frameworks. |
| Energy and safety design records | ASHRAE 90.4-2025; ASHRAE 15/34; NFPA 70, 75 and 110 as applicable | Design-point heat/power, native heat-balance record, distribution and transfer assumptions | Annual infrastructure energy, refrigerant/charge, electrical protection and emergency-system classification require the selected equipment and site design basis. |
| Jurisdiction and editions | Adopted building, mechanical, fire, plumbing, electrical and energy codes | Concept engineering basis; no construction site assigned | The construction design establishes adoption dates, amendments, authority having jurisdiction and applicable clauses. |
Standards, equipment coordination and evidence register
Design-basis selection
Jurisdiction and adopted codes
Building, mechanical, plumbing, fire, electrical and energy codes ↗A project location establishes the applicable editions, local amendments and authority having jurisdiction. Published editions and adopted requirements are recorded separately.
- Evidence to inspect
- Location, adoption record, occupancy/use, equipment arrangement and review of applicable clauses.
Selected-condition calculations
Equipment environmental conditions
ASHRAE TC 9.9 guidance + equipment manufacturer limits ↗Air inlet, humidity, dew point and liquid temperature ranges are selected for the actual IT equipment class. Recommended and allowable conditions serve different purposes.
- Evidence to inspect
- Rack-inlet conditions, fluid properties and temperatures, selected equipment limits and dew-point margin.
Energy-method reference
Infrastructure energy
ANSI/ASHRAE 90.4 ↗The energy assessment addresses the relevant mechanical and electrical infrastructure methods in the selected edition. A single design-point power balance is one input to that assessment.
- Evidence to inspect
- Applicable edition, weather/load profiles, part-load curves, component energy and electrical-loss boundary.
Equipment-specific review
Refrigeration safety
ASHRAE 15 and 34 + adopted mechanical provisions ↗DX and chiller selections connect refrigerant safety classification to system location, charge, ventilation and the applicable protection measures.
- Evidence to inspect
- Refrigerant, charge, machinery-room/occupied-space arrangement and equipment documentation.
Coordination reference
Fire and electrical coordination
NFPA 75 + applicable adopted fire/electrical requirements ↗Cooling pipework, immersion equipment, electrical supplies, detection and shutdown sequences must be coordinated with fire protection and maintenance access.
- Evidence to inspect
- Equipment listing/documentation, fluid characteristics, fire strategy and power/control diagrams.
Topology scenario checks
Resilience and maintainability
Uptime Tier criteria or ANSI/TIA-942-C rating framework ↗Capacity redundancy, distribution paths, shared dependencies and maintenance states are evaluated together. N+1 and 2N describe arrangements; a Tier or rating requires assessment of the complete infrastructure.
- Evidence to inspect
- A/B paths, isolation points, power dependencies, failure scenarios and integrated testing.
Component-selection reference
Liquid-system integration
OCP cooling specifications + manufacturer requirements ↗Fluid compatibility, cleanliness, connections, filtration, leak detection and CDU interfaces are part of specifying a working liquid-cooling installation.
- Evidence to inspect
- Fluid specification, material compatibility, loop pressure/flow, connection and maintenance requirements.
Reference register · reviewed 28 October 2026
ASHRAE lists Standard 90.4-2025 as its published data-center energy standard. TIA lists revision C, May 2024. The applicable project editions and amendments are established through jurisdictional and contractual review.
- NFPA 70 · National Electrical Code
- NFPA 110 · Emergency and Standby Power Systems
- ASHRAE · Data Centers and Telecommunication Facilities ↗
- DOE · Energy-efficient data center design ↗
- Uptime Institute · Tier classifications ↗
- TIA · ANSI/TIA-942-C ↗
- ASHRAE · Standards titles, purposes and scopes ↗
- NFPA · Standard 75 ↗
- Open Compute Project · Cooling Environments ↗
- ICC · Model codes and standards ↗
Results and engineering decisions
The results distinguish three design decisions. Containment improves rack-inlet conditions by limiting recirculation. Redundant capacity and distribution determine how much heat can still be removed during a fault. Cooling backup power and coupled thermal inventory determine whether circulation and supply temperature remain controlled during generator transfer.
In the default 192 kW comparison, N uses three 80 kW duty paths. One module failure leaves 160 kW for 240 seconds, accumulating 2.133 kWh of cooling shortfall. N+1 restores the duty after a 30-second pickup, reducing shortfall to 0.267 kWh. The same single-failure result for N+2 reflects an unused second spare; its value appears when a component is already under maintenance. These results demonstrate why capacity reserve does not guarantee continuity. In the 85%-load airflow reference, the approximately 3.0°C containment benefit is an inlet-mixing result, while the 23.66°C unit-outage peak includes the timed thermal response. In Case B, maintaining both coolant loops and residual-air cooling through the 60-second bridge avoids the flow interruption present without cooling UPS support. Inspect the comparison below for the calculated peak, energy requirement and sensitivity to inventory. Every result remains attached to its named case and explicit assumptions.
Reference results and their design consequence
Airflow reference · 24 racks · 85% load · 163.2 kW
At the reference load, the selected recirculation assumptions lower the hottest inlet by approximately 3.0°C (23.40°C open versus 20.43°C contained). The contained unit-outage trace peaks at 23.66°C. This analytical mixing sensitivity uses 25% open-aisle and 4% contained recirculation; the main sizing lab is a separate 100%-operating-load case.
Transfer reference · one NVL72 rack · 330 kW
With both loops and air cooling backed, the model retains flow and reaches 41.8°C peak flowing TCS supply during the 5-minute window. The 60-second bridge requires 0.19 kWh nominal battery energy at the entered usable fraction and conversion efficiency.
| Cooling backup | No TCS flow · s | Peak flowing TCS · °C | Liquid / air heat unserved · kWh |
|---|---|---|---|
| No cooling UPS | 60.0 | Interrupted: 60.0 s without flow. 40.00°C while flowing only | 5.390 / 0.110 |
| Technology pump only | 0.0 | 58.68°C | 0.000 / 0.110 |
| Both loops + residual air | 0.0 | 41.85°C | 0.000 / 0.000 |
The coupled model solves approximately 596 L of facility volume for the 45°C TCS reference over this window, assuming uninterrupted power to both loops and the air unit. This value depends on the 250 L technology inventory, effective CDU conductance, initial temperatures and recovery control. The 2,000 L reference buffer is evaluated with those same assumptions.
Sensitivity to connected buffer volume and CDU coupling
These half-to-double ranges bracket the selected 2,000 L participating facility inventory and 65 kW/(K·rack) effective CDU coupling. They illustrate model sensitivity, not measured equipment variation. Every row keeps one 330 kW rack, 250 L technology inventory, 40°C initial TCS supply, a 60 s transfer, 170 kW recovery margin and the 1 kWh both-loop/air UPS. Initial facility temperature is recalculated consistently with each conductance.
| Facility inventory · L | CDU coupling · kW/(K·rack) | Peak flowing TCS · °C | Above 45°C · s | No flow · s |
|---|---|---|---|---|
| 1000 | 32.5 | 43.0 | 0.0 | 0.0 |
| 1000 | 65 | 43.3 | 0.0 | 0.0 |
| 1000 | 130 | 43.6 | 0.0 | 0.0 |
| 2000 | 32.5 | 41.8 | 0.0 | 0.0 |
| 2000 | 65 | 41.8 | 0.0 | 0.0 |
| 2000 | 130 | 42.0 | 0.0 | 0.0 |
| 4000 | 32.5 | 41.0 | 0.0 | 0.0 |
| 4000 | 65 | 41.0 | 0.0 | 0.0 |
| 4000 | 130 | 41.0 | 0.0 | 0.0 |
The resulting design sequence is to control recirculation, size the liquid and air paths independently, verify the response of every serial component, and test cooling power continuity with finite battery energy. A sustainable IT capacity, an actual heat-removal rate and an IT shutdown are distinct results throughout the calculations.
