With the cooling stopped and the servers still running on UPS, the air in a mid-size hall can climb about 5.5 °F (3 °C) every 6 seconds if nothing else absorbs the heat. How long a hall lasts depends on what else is absorbing heat, and on what stays running while the chillers are down. This article does the arithmetic, then says what a steady-state model can and cannot tell you.
The air stores almost nothing
A cubic meter of air at room conditions holds about 1.21 kJ of heat for each degree of temperature rise (density 1.204 kg/m³ times a specific heat of 1.006 kJ/kg·K). A cubic meter of water holds about 4,180 kJ per degree. Water stores roughly 3,400 times more heat than the same volume of air, so one liter of water has the heat capacity of about 3.4 m³ of air.
Take the room in our transient CRAC failure white paper: 1,800 ft² (167 m²) with a 10 ft (3.05 m) ceiling and 315.25 kW of IT load. Its air volume is 510 m³, so the air holds 617 kJ per kelvin. If the air alone absorbed the whole load, the temperature would rise at
That is 55 °F (31 K) per minute. The air would pass the 9 °F (5 K) gap between ASHRAE’s recommended 80.6 °F (27 °C) and the A1 allowable 89.6 °F (32 °C) in 10 seconds. Add the 18 in (0.46 m) raised-floor plenum to the volume, which is 76 m³ more, and the rate falls only to 0.44 K/s, 11 seconds for the same gap.
This is a bound, not a prediction. The racks, the floor and the building also absorb some heat, which slows the real rise, and air does not mix evenly, so some intakes pass a limit before the room average does. It explains why the studies below measure seconds. In a model hall of 234 racks and 936 kW that we stopped for a failure test, the first rack passed 27 °C after 14.5 seconds, as the cooling-failure studies page shows.
What actually buys time
Fans on UPS. Servers move air through the room whether or not the cooling units run. If the cooling unit fans stop, the air that servers push through the stopped units is not cooled, and the hot exhaust recirculates to the intakes. In one of our model halls with 18 units, the units passed about 60% of their normal airflow with their fans stopped, all of it uncooled. Fans on UPS keep the air circulating through the coils, which matters only if the coils are still cold.
Chilled water in the loop. The pipes, coils and any buffer tank hold cold water. Take 20,000 L of chilled water that can warm from 44.6 °F to 59 °F (7 °C to 15 °C) before it stops cooling the air usefully. That is an assumed example, not a typical hall:
At 315 kW that is 2,120 seconds (35 minutes). At 936 kW it is 714 seconds (12 minutes). The stored water only works if pumps and unit fans keep running to bring it to the air. Without them it is cold water in a pipe, and the room sees none of it.
What sits on UPS therefore decides the curve. Three common cases, from worst to best:
| On UPS | What the room sees |
|---|---|
| Servers only | Heat keeps coming, no air moves through the coils, and the room follows the air-only arithmetic above |
| Servers and cooling unit fans | Air circulates, but stored cold only helps if the water is still moving |
| Servers, fans and pumps | Stored water in the loop is available until it warms or until the chillers return |
Chiller restart. When utility power drops, the generators start and the chillers have to restart and ramp. The restart times we were given in delivered studies ran from 60 to 120 seconds. In the first minute or two the water in the loop is the only cooling there is.
Why a steady model cannot answer this
A steady-state model gives the temperatures once everything has stopped changing. It assumes the cooling runs. Remove the cooling and the model has no steady answer: the temperature keeps rising until something else changes.
“How long” is a question about time, so it needs the equations solved over time. A transient study starts from the settled steady-state room, applies the failure sequence (what stops, what stays on UPS, when each piece restarts and how fast it ramps), and records the air temperature at each rack intake every fraction of a second. The result is a curve per rack that can be read against the ASHRAE limits: 80.6 °F (27 °C) recommended and, for class A1 equipment, 89.6 °F (32 °C) allowable.
What ASHRAE says about rate of change
ASHRAE’s 2021 thermal guidelines reference card (Fifth Edition, TC 9.9) limits the temperature change of most IT equipment to 36 °F (20 °C) in an hour and no more than 9 °F (5 °C) in any 15-minute period. Tape storage gets a stricter 9 °F (5 °C) in an hour. ASHRAE describes these as temperature changes within a window, not as rates.
Those limits were written for slow drift in a running facility, and they put a number on how fast the intake air may change. Against the arithmetic above, a 5 K change in about 10 seconds is roughly 90 times faster than the 15-minute limit allows. Read that as our interpretation: the guideline does not address a utility failure, but it shows how far outside normal operation the first minute of a cooling loss is. Whether a given server rides through depends on its own specification, which the manufacturer has to confirm.
What the studies have covered
Since June 2023 our engineers have delivered 18 cooling-failure studies for 10 customers. The scenarios customers asked for fall into five types:
- All cooling lost, with IT on UPS.
- Cooling unit fans on UPS, nothing else.
- Fans and pumps on UPS, riding through on the water already in the loop.
- One unit, or some of the units, lost.
- Full power loss, with the chillers back after a restart delay.
The sequence assumed (what is on UPS, when each unit restarts) moves the answer as much as the room does, so we agree it with the customer before anything is run.
A transient study is an engineering service
CoolSim solves steady state. The time-dependent study is an engineering service: our engineers run it from your CoolSim model and your failure sequence, and send you the results.
To scope one, we need the room model, what is on UPS, when generators pick up and when each fan, pump and chiller restarts, and the IT load during the event. The cooling-failure studies page lists what a study delivers, with a worked example.