Cooling efficiency and free cooling

Cooling is the largest source of infrastructure overhead in most data centers, and the largest opportunity for savings. A facility that cools with mechanical refrigeration all year at low temperatures may spend 40 % or more of its IT energy on cooling; one that runs warm, contains its airflow and uses outdoor air or water for most of the year may spend 5 to 15 %. The physics of cooling efficiency is straightforward: move heat with as little fan and pump energy as possible, run compressors as little as possible and with as small a temperature lift as possible, and use the outdoor environment whenever it is cold enough. This lesson covers where cooling energy goes, the effect of temperature setpoints, airflow management, economizers and free cooling, chiller efficiency, liquid cooling, refrigerants, and the energy codes that apply. The mechanical design of these systems is covered in DCE-330; here the focus is efficiency.

Where cooling energy goes

Cooling energy is consumed by:

• Compressors in chillers or direct-expansion units, which pump heat from a low temperature to a higher one. They are usually the largest consumer when they run.
• Fans in air handlers, computer room units and fan walls, which move air through the IT equipment and coils.
• Pumps that circulate chilled water, condenser water or coolant.
• Heat rejection equipment: cooling tower fans, dry cooler fans and, in evaporative systems, water pumping and treatment.
• Humidification and dehumidification, where used.

Each can be reduced by different means: compressors by raising temperatures and using free cooling; fans and pumps by variable speed drives, lower pressure drops and better airflow management; heat rejection by larger, more efficient equipment.

Temperature setpoints

For decades, data halls were kept cold, around 18 to 21 °C, out of caution. ASHRAE's thermal guidelines for data processing environments, developed with IT manufacturers, recommend server inlet temperatures from 18 to 27 °C for most equipment, with allowable ranges extending to 32 °C (class A1) and beyond for more tolerant classes. Raising supply air and chilled water temperatures within these ranges has two major benefits. First, chillers work across a smaller temperature lift and become more efficient. Second, the outdoor temperature below which free cooling is possible rises, so economizers run for many more hours. The trade-offs are that server fans may speed up at higher inlet temperatures, consuming some of the savings, and that ride-through time after a cooling failure shortens, because there is less margin before equipment reaches its limits (DCE-450, Lesson 4). The optimum is usually found by measuring total facility energy, including server fan energy, at different setpoints.

Airflow management

Air cooling is efficient only if cold air goes through the IT equipment and hot air returns to the coolers without mixing. Mixing forces the cooling system to supply colder air or more of it than necessary. Good airflow management includes hot aisle or cold aisle containment, blanking panels in empty rack spaces, sealing cable openings and gaps under racks, and matching the airflow supplied by coolers to the airflow drawn by IT equipment. Once airflow is controlled, supply temperatures can be raised and fan speeds lowered.

Fan speed matters enormously because of the fan affinity laws: airflow is proportional to fan speed, pressure to the square of speed, and power to the cube of speed. Small reductions in speed produce large savings:

Running eight fans at 50 % speed uses far less power than running four at 100 %, even though they move the same air, which is why redundant fans are best operated together at reduced speed rather than with standby units switched off.

Economizers and free cooling

Free cooling uses the outdoor environment to remove heat without running compressors. The main approaches are:

• Direct air-side economizers: outdoor air is filtered and brought directly into the data hall, with hot air exhausted. Very efficient, but requires good filtration and humidity control, and is sensitive to smoke and pollution events.
• Indirect air-side economizers: outdoor air cools the data hall air through a heat exchanger without mixing, often with evaporative assistance on hot days. They avoid contamination at some cost in efficiency.
• Water-side economizers: cooling towers or dry coolers cool the chilled water directly through a heat exchanger when outdoor conditions allow, bypassing the chillers. Partial economization pre-cools the water so that chillers work less.
• Evaporative cooling: water evaporation lowers air or water temperature toward the outdoor wet-bulb temperature, greatly extending free cooling in dry climates at the cost of water consumption (Lesson 6).

How many hours of free cooling a site gets depends on its climate and on how warm the cooling water or air is allowed to be.

import numpy as np

# Hourly outdoor dry-bulb temperature for a year: seasonal and daily cycles plus weather noise
h = np.arange(8760)
rng = np.random.default_rng(4)
def climate(mean, seasonal, daily):
    return (mean - seasonal * np.cos(2 * np.pi * (h / 8760 - 0.03))
            - daily * np.cos(2 * np.pi * (h % 24 - 3) / 24) + rng.normal(0, 3, h.size))
sites = {"Cool site (mean 6 C)": climate(6, 14, 5), "Warm site (mean 20 C)": climate(20, 8, 6)}

approach = 5.0   # dry cooler approach: water leaves this much warmer than outdoor air
print("Hours of full free cooling with dry coolers (approach 5 K)")
for name, t in sites.items():
    for supply in (12, 18, 24):
        full = (t <= supply - approach).mean()
        partial = ((t > supply - approach) & (t < supply + 2)).mean()
        print(f"  {name:<22} supply {supply} C: full {full:5.1%}, partial {partial:5.1%}")

# Fan affinity law: power scales with the cube of speed
for speed in (1.0, 0.9, 0.8, 0.6):
    print(f"Fan at {speed:.0%} speed: {speed**3:5.1%} of full power, {speed:.0%} of airflow")

Hours of full free cooling with dry coolers (approach 5 K)
  Cool site (mean 6 C)   supply 12 C: full 52.6%, partial 18.6%
  Cool site (mean 6 C)   supply 18 C: full 68.6%, partial 19.9%
  Cool site (mean 6 C)   supply 24 C: full 85.8%, partial 12.6%
  Warm site (mean 20 C)  supply 12 C: full  4.8%, partial 18.4%
  Warm site (mean 20 C)  supply 18 C: full 19.9%, partial 29.8%
  Warm site (mean 20 C)  supply 24 C: full 45.3%, partial 31.3%
Fan at 100% speed: 100.0% of full power, 100% of airflow
Fan at 90% speed: 72.9% of full power, 90% of airflow
Fan at 80% speed: 51.2% of full power, 80% of airflow
Fan at 60% speed: 21.6% of full power, 60% of airflow

At the cool site, raising the supply water from 12 °C to 24 °C raises full free cooling from about half the year to over 85 %, with partial economization most of the remaining hours. At the warm site, free cooling is rare at 12 °C but covers almost half the year at 24 °C. The same physics explains why liquid-cooled AI clusters, which can accept coolant at 30 °C or warmer, can run on dry coolers almost everywhere for most of the year. The fan lines show that a fan at 80 % speed uses only about half the power of one at full speed.

Chiller efficiency

Chiller efficiency is expressed as the coefficient of performance (COP), the heat removed divided by the electrical energy used, or in North America often as kilowatts per ton of refrigeration (kW/ton, where one ton is 3.517 kW of cooling). A COP of 6 corresponds to about 0.59 kW/ton. COP depends strongly on the temperature lift between the evaporator (chilled water) and condenser (heat rejection) sides. The theoretical maximum, the Carnot COP, shows the trend:

with temperatures in kelvins. Real chillers achieve perhaps half of the Carnot value, but they follow the same pattern.

Other chiller efficiency measures include variable speed compressors, which are much more efficient at part load; magnetic bearing centrifugal compressors; selecting chillers by their integrated part load value (IPLV), which reflects efficiency at the loads they actually run at, rather than full-load efficiency alone; and sequencing chillers so that each runs near its most efficient load.

Liquid cooling

Direct-to-chip and immersion cooling carry heat away in liquid, which holds far more heat per unit volume than air. Because liquid cooling can accept warm coolant, often 30 to 45 °C, it can reject heat with dry coolers for most or all of the year without chillers, and it reduces or eliminates server and air handler fans. It also produces heat at temperatures more useful for reuse (Lesson 9). The trade-offs are cost, the need for coolant distribution infrastructure, and the fact that some heat is still rejected to air in the room, so hybrid air and liquid systems are common. For high-density AI clusters, liquid cooling is often the only practical option, and it can achieve partial PUE values for cooling of a few percent.

Refrigerants

Chillers and direct-expansion units use refrigerants, many of which are hydrofluorocarbons (HFCs) with global warming potentials thousands of times that of CO₂. Leaks are Scope 1 emissions (Lesson 7) and can be significant. International agreements (the Kigali Amendment to the Montreal Protocol) and national rules are phasing down HFCs, and new equipment increasingly uses low-GWP refrigerants such as HFOs, ammonia, CO₂ or hydrocarbons, each with its own safety considerations. Leak detection, maintenance and responsible recovery at end of life matter as much as the choice of refrigerant.

Energy codes for data center cooling

In the United States, ASHRAE 90.4, Energy Standard for Data Centers, adopted by reference in many energy codes, sets maximum values for a design mechanical load component (MLC), representing cooling energy relative to IT load by climate zone, and a maximum electrical loss component (ELC), giving designers flexibility in how they achieve compliance.

Codes set minimums; leading designs go far beyond them, and owners increasingly specify target PUE, WUE and partial PUE values in their design briefs.