Data center cooling system types and how to select one

Data center cooling system types, and the axes that pick one

A data center cooling system is the equipment chain that removes the heat the IT load makes and rejects it outside, and the type you pick comes down to three axes. The first is the medium that carries the heat off the equipment: air or liquid. The second is where you capture the heat: at the room, the row, the rack, or the chip. The third is how you finally reject it outdoors: to air or to water. Name those three for a given hall and you have named the system.

Every architecture in this guide is a different combination of those three choices. A small office server room is room-level air capture with refrigerant rejection. A large enterprise hall is room or row air capture fed by a chilled water plant rejecting to a cooling tower. An AI hall is chip-level liquid capture with air handling the rest, rejecting to a tower or a dry cooler. Same job, different points on the same three axes.

This guide is the selection hub. It maps the families and the tradeoffs and points to the deep guides where the detail lives. The airflow and containment work that makes air cooling carry its load is covered in the cooling and airflow guide, and the cold-plate liquid loop that AI racks need is covered in the direct-to-chip guide. Read those for the inside of the box. Read this to decide which box.

The heat path every system follows

Every cooling system, no matter how it is built, moves heat along the same four-step chain: the chip makes it, a medium carries it away, a cooling unit picks it up, and a heat-rejection device dumps it outside. Trace that chain and you can place any product on the market, because each type is just a different way of doing one or more of those steps.

Start at the silicon. Nearly all the power a server draws turns into heat, so a 15 kW rack is a 15 kW heater that never shuts off. Step one is getting that heat off the components, with air blown across heatsinks or coolant run through a cold plate. Step two carries it to a cooling unit: air to a CRAC or CRAH coil, or liquid to a coolant distribution unit. Step three is the cooling unit handing the heat to a water or refrigerant loop. Step four rejects it to the outdoors at a cooling tower, a dry cooler, or a condenser.

The reason to hold the whole chain in your head is that a system is only as good as its weakest step. A perfect cold plate on the chip does nothing if the facility water loop behind it cannot reject the heat. A big chiller plant does nothing if the air never reaches the rack inlet. When a hall runs hot, the fault is somewhere on this chain, and knowing the steps tells you where to look.

Air cooling or liquid cooling?

The first real split is air versus liquid, and density decides it. Air cooling is the traditional approach: fans push cold air through the racks and carry the heat to a coil. It works, it is well understood, and it cools the large majority of the installed base, but it has a ceiling. Liquid cooling runs water or a dielectric fluid to the heat instead of air, and it exists because air ran out of room at high density.

The physics is the reason. A given volume of water carries thousands of times more heat than the same volume of air, so liquid moves the same kilowatts in a fraction of the flow and gets the cooling within inches of the chip. Below the air ceiling, air is cheaper and simpler and there is no reason to plumb liquid. Above it, no amount of cold air solves the problem, because you cannot physically push enough air through a dense rack to carry the heat.

The market is moving with the density. The liquid cooling segment has been growing fast as AI loads land, with a large share of facilities still air-only but most planning liquid within a few years. The honest framing is not that liquid replaces air. It is that the high-density end moved to liquid and is not moving back, while everything below the ceiling stays on air. The direct-to-chip guide covers the liquid side in depth.

What is the difference between a CRAC and a CRAH?

A CRAC is a computer room air conditioner with its own refrigerant circuit, a direct-expansion unit that cools the air with a compressor and rejects heat to a condenser outside. A CRAH is a computer room air handler with no compressor; it is a coil fed with chilled water from a central plant, and a valve modulates the water flow to hold supply temperature. Both are the air-capture workhorses. The difference is where the cold comes from.

The choice tracks scale. CRAC and packaged DX units suit smaller halls and edge sites where a central chilled water plant is not worth the capital, because each unit makes its own cold. CRAH units dominate large facilities, because one efficient central chiller plant feeding many air handlers beats a field of individual compressors on energy and maintenance. The trap on submittals is that the trade uses CRAC loosely for both. Read the cut sheet: a compressor in the unit makes it a CRAC, a chilled water coil makes it a CRAH.

Both are room or perimeter machines in their classic form, and both follow the same air-path rules. How the air actually reaches the rack inlet, through a raised floor or overhead, and the containment that keeps it from mixing, is the airflow side covered by topic in the cooling and airflow guide.

Room-level cooling: perimeter units and the raised floor

Room-level cooling is the traditional approach where perimeter CRAC or CRAH units cool the whole hall as one volume, and it is where most legacy data centers started. The units sit around the edge of the room and blow cold air into the space; the racks draw it in, exhaust it hot, and the hot air finds its way back to the units. It is simple, it is cheap to build, and it works as long as the per-rack density stays low.

The classic arrangement is downflow units pushing cold air into a sealed raised-floor plenum, with the plenum pressure pushing air up through perforated tiles to the cold aisle and the server inlets. The return rises overhead and loops back to the units. Newer halls often skip the raised floor and supply overhead, but the principle is the same: the whole room is the distribution path.

Room cooling runs out of headroom because the air has a long way to travel and plenty of chances to mix before it reaches the rack. Once racks climb past roughly 10 to 15 kW, the cold supply tends to blend with hot exhaust before it gets to the inlet, and the far racks starve. That density limit is what pushed the industry toward capturing heat closer to the load. The raised-floor air distribution detail lives in the airflow guide.

Hot aisle, cold aisle, and containment

Hot-aisle/cold-aisle is the layout that makes room and row air cooling work, and containment is the barrier that closes it. Racks face each other so cold supply enters one aisle and hot exhaust dumps into the next, and containment, doors at the aisle ends and a roof over it, stops the two air streams from mixing. It is the single biggest air-side efficiency move there is, and it costs panels and doors, not megawatts.

Without it, cold supply short-circuits to the return without passing through a server (bypass), and hot exhaust loops back into the cold aisle and raises inlet temperatures (recirculation). Operators then drop the supply temperature for the whole room to satisfy the worst rack, which over-cools everything else and wrecks efficiency. Containment lets the cold aisle stay cold end to end so you can run warmer supply and the plant does less work.

This guide keeps containment to the headline because it is air-management detail, not a system type. The full treatment, including how the airflow, delta-T, and rack power lock together and how pressure proves the containment is doing its job, is in the cooling and airflow guide.

Row-level and in-row cooling

In-row cooling moves the cooling unit out of the perimeter and into the line of racks, so the air path shrinks from across the room to a few feet. The units sit between the racks, pull hot exhaust straight from the hot aisle, cool it, and discharge cold supply right at the cold aisle in front of the server inlets. Shortening the path is the whole idea: less distance means less mixing and tighter control of the inlet temperature.

This is the step up from room cooling for mid-density work, commonly in the range above 15 kW where room air starts to lose the inlet. Because each unit serves the racks next to it, you can match cooling to the load row by row instead of conditioning the entire hall to its hottest spot. In-row units are also called close-coupled cooling, because the cooling is coupled close to the heat it serves.

In-row units come in chilled water and DX versions, so the choice of central plant or self-contained refrigerant still applies. They lean hard on containment to work, because a short air path only helps if the supply and return are kept apart. Push the density high enough and even in-row air becomes marginal, which is where the heat capture moves onto the rack itself.

Rack-level cooling: the rear-door heat exchanger

A rear-door heat exchanger captures the heat at the rack by replacing the cabinet's back door with a liquid-cooled coil. The server fans push hot exhaust through the coil on the way out, the coil pulls the heat into a water loop, and the air that enters the room is already cooled, sometimes back to room neutral. It is the first member of the liquid family most halls meet, because it is air-to-liquid at the rack without rebuilding the room.

Rear-door units are the bridge between air and full liquid cooling. They can take a hall well past what room or in-row air handles, into the range around 30 to 50 kW per rack, and a well-designed door can capture nearly all of a rack's heat. Because the servers themselves are unchanged, still air-cooled inside, a rear-door retrofit is the lowest-disruption way to add liquid capacity to an existing air-cooled hall.

The catch is that a rear-door needs a water loop and the controls that go with it, so it brings the leak-detection and flow-balancing discipline of liquid cooling to a rack that used to be pure air. Active doors with their own fans also draw power and need their controls commissioned. The rear-door sits in the same liquid-cooling family as direct-to-chip and immersion, covered by topic in the direct-to-chip guide.

Direct-to-chip: the cold plate on the chip

Direct-to-chip cooling clamps a cold plate straight onto the CPU or GPU and runs coolant through it to pull heat off at the source. It is the mainstream choice for AI racks, because it bolts onto a fairly conventional server, reuses much of the existing air handling, and scales to the densities the GPUs need. The cold plate takes the hottest chips, commonly the 70 to 80 percent of the rack heat the silicon makes, and air still handles the memory, drives, and power supplies.

That hybrid nature is the part people miss. A direct-to-chip rack is not a sealed liquid box; it is liquid on the chips and air on the rest, so the hall still needs air handling sized for the residual load. The coolant runs through a clean technology loop that a coolant distribution unit isolates from the building water, keeping the dirty facility side away from the silicon.

This guide keeps direct-to-chip brief because it has its own deep guide. The cold plate, the two loops, single-phase versus two-phase, the coolant chemistry, and the leak strategy are all covered there. The point for selection is simple: when the per-rack density crosses what air can hold, direct-to-chip is the form most of the market took.

Immersion cooling: the server in the fluid

Immersion cooling drops the whole server into a bath of dielectric fluid, so the liquid touches everything and carries essentially all the heat. There are no server fans, because there is no air path inside the tank. It is the densest of the liquid options and the most efficient, and it reaches densities that direct-to-chip and rear-door cannot, which is why it keeps coming up for the most extreme AI and HPC loads.

Immersion runs in two flavors. Single-phase keeps the fluid liquid and pumps it through a heat exchanger; two-phase lets the fluid boil on the hot components and carry heat away as latent heat of vaporization, which moves more heat at lower flow. ASHRAE's liquid-cooling work treats both, and two-phase is the one to watch as rack power climbs past what single-phase comfortably holds, though it brings vapor and pressure management and costlier fluids.

The reasons immersion is not everywhere are physical and practical. The tanks are heavy, so floor loading is a real constraint; the dielectric fluid has to be stored, handled, and kept clean; and retrofitting a tank-based layout into a hall built for racks is the hardest of the liquid options. Where the weight and the fluid handling are acceptable, immersion is the efficiency and density leader. The liquid family it belongs to is covered by topic in the direct-to-chip guide.

The chilled-water plant behind the air and liquid

Behind most CRAH units and most liquid loops is a chilled-water plant, the central system that makes cold water and moves it where the heat is. Chillers cool the water, traditionally in the low 40s F and warmer on efficient and free-cooling designs, and pumps push it through a piping loop to the cooling coils. The coil pulls heat from the air or the technical loop into the water, the warmed water returns to the chiller, and the chiller rejects that heat outside.

The plant is the shared engine that several of the system types plug into. A CRAH is a chilled-water coil. A rear-door is a chilled-water or facility-water coil. A direct-to-chip CDU rejects into facility water that a chilled-water plant or a tower supplies. So the plant choice ripples through every air handler and liquid loop downstream of it, which is why a large facility builds one efficient central plant rather than a field of standalone units.

The flip side is that the plant is a chain of single systems that all have to hold: chillers, pumps, valves, and the piping loop. Before it carries load, the loop gets a hydrostatic pressure test and a flush, and its redundancy and ride-through have to be proven under load. The plant itself, with its cooling tower and chiller, is a large subject covered by topic in the cooling and airflow guide.

DX and packaged refrigerant cooling

Direct-expansion cooling, DX, makes the cold with a refrigerant circuit right at the unit instead of pulling chilled water from a central plant. A DX CRAC has a compressor and an evaporator coil inside, and it rejects its heat through refrigerant piping to a condenser or condensing unit outside. Packaged DX shows up as the perimeter CRAC, as self-contained in-row units, and as the small wall-mount or ceiling units in a closet-sized server room.

DX earns its place where a chilled-water plant is not justified. On a small hall, an edge site, or a single equipment room, building and maintaining a chiller plant costs more than the load is worth, so a packaged refrigerant unit that makes its own cold is the practical answer. The tradeoff is efficiency at scale: a yard full of individual compressors uses more energy than one large central chiller doing the same total work.

DX also limits the free-cooling story. A refrigerant unit can carry an economizer, but the central chilled-water plant is where waterside and airside economizing pay off most. For a small or distributed footprint, DX is the right tool. For a large hall chasing a low PUE, the chilled-water plant usually wins.

How does the heat finally leave the building?

The heat leaves the building at the heat-rejection device, the last step in the chain, and the four common choices trade water against energy against climate. A cooling tower rejects heat by evaporating water, which is efficient but consumes water. A dry cooler blows ambient air over a coil with no evaporation, which uses no process water but more fan energy and depends on the outdoor temperature. A chiller uses a compressor to lift the heat to a temperature it can shed in any climate, at the highest energy cost. An air-cooled condenser rejects DX refrigerant heat straight to outdoor air.

Most large plants combine them. A water-cooled chiller plant rejects through a cooling tower; an air-cooled chiller plant rejects through its own condensers; a warm-water liquid loop may reject through a dry cooler or a tower running in free-cooling mode for much of the year, with a chiller only as backup for the hottest hours. The warmer the loop you can run, the more the cheap devices carry the load and the chiller sits off.

The selection question is which device fits the site. Water-scarce or expensive-water sites lean toward dry coolers and air rejection. Hot climates lean on the chiller more hours. The cooling-tower and chiller detail is covered by topic in the cooling and airflow guide; the point here is that this last step is where the water-versus-energy decision is actually made.

Free cooling and economizers

Free cooling is the efficiency play that uses cold outdoor air or water to carry the load when the weather allows, idling the compressors that otherwise dominate the energy bill. It comes in two forms. An airside economizer brings filtered outside air straight into the hall instead of running the refrigeration. A waterside economizer makes chilled water with the cooling tower or dry cooler alone, letting the chiller compressors sit off while the outdoor conditions do the work.

The savings are real and large. Depending on the climate and the supply temperature, economizing can cut the precision-cooling energy on the order of 30 to 50 percent, and in cold or dry climates the plant can run free for much of the year. This is why warmer supply temperatures and wider ASHRAE envelopes matter so much: every degree warmer you can run the hall is more hours the economizer carries the load.

Free cooling is a system-selection input, not just a control setting. A plant designed for waterside economizing needs the tower capacity and the heat exchanger to do it; an airside design needs the louvers, filtration, and humidity control to bring outside air in safely. Build the economizer in from the start, because retrofitting free cooling into a plant that was not laid out for it is expensive. The economizer changeover logic is covered by topic in the cooling and airflow guide.

What is the most efficient data center cooling?

The most efficient cooling is the one that rejects the most heat for the least overhead energy, and on the metric the industry uses, that means the lowest PUE. PUE, power usage effectiveness, is total facility energy divided by the energy that reaches the IT equipment; 1.0 is the unreachable floor, older halls run near 2.0, and good modern designs land around 1.2 to 1.4. Cooling is where most of the non-IT energy lives, so the cooling choice drives the PUE more than any other system.

In practice the efficient end is a warm-running plant with containment, high delta-T, and as many free-cooling hours as the climate allows, rejecting through evaporative or dry cooling rather than leaning on the chiller. Liquid cooling helps at high density because it moves heat with far less fan energy and tolerates warmer supply, opening more free cooling. There is no single most-efficient product; there is the combination that fits the density and the climate.

Two cautions keep this honest. First, a low PUE bought with an evaporative tower can come at a high water cost, so PUE alone does not tell the whole efficiency story, which is the WUE tradeoff below. Second, ASHRAE Standard 90.4 is the code-side companion that bounds the mechanical and electrical overhead by climate zone rather than regulating PUE directly. PUE is the operating scorecard; 90.4 is the design path. The PUE detail is covered by topic in the cooling and airflow guide.

Water versus air heat rejection, and the WUE tradeoff

The water-versus-air rejection choice is a direct tradeoff: evaporative cooling uses water and less energy, while dry and air-cooled rejection uses no process water and more energy. A cooling tower evaporates water to shed heat and can let a water-cooled plant draw meaningfully less electricity than an air-cooled one, especially on hot afternoons. A dry cooler skips the water entirely but works the fans harder and is limited by how cold the outside air gets.

WUE, water usage effectiveness, is the metric that puts numbers on it, measured in liters per kilowatt-hour of IT energy. Many evaporative sites land near the industry average around 1.8 to 1.9 L/kWh, while best-in-class and dry-cooled designs push toward 0.3 to 0.7 or lower. The reason WUE has moved up the priority list is that a large share of the water a tower uses evaporates and never returns to the supply, which matters where water is scarce or contested.

This is where optimizing PUE alone goes wrong. Chase the lowest PUE with an evaporative tower in a water-stressed region and you trade an energy win for a water problem. The 2026 direction is visible in operators piloting closed-loop, zero-water-evaporation designs that accept a slightly higher PUE to drop WUE toward zero. Pick the rejection method against both numbers and the site's water reality, not PUE by itself.

How rack density decides the system

Per-rack power is the single strongest input to the cooling-system choice, and it sorts the families into a rough ladder. Low density is room air. Mid density is containment and in-row. High density is rear-door or liquid at the rack. Very high density, the AI and HPC end, is direct-to-chip or immersion. Knowing roughly where a hall sits on that ladder tells you which family is even in the running before any other factor.

The bands are practical ranges, not hard lines, and they shift with the containment, the supply temperature, and how much fan energy you will spend. Air with good containment can be pushed surprisingly far, and a rear-door buys headroom at the rack before you commit to plumbing the servers. But the trend is one-way: AI accelerators that ran a few hundred watts now run past a thousand, and a single dense AI rack can pull past 100 kW, a range that was never an air problem to begin with.

The expensive mistake is designing for today's density and ignoring the next generation. A hall built for 10 to 20 kW racks cannot absorb a 100 kW AI row without a liquid path, and retrofitting one in is harder than planning it. Size the cooling for the density the hall will actually carry, including the load the tenant has not told you about yet. The density-to-system map is the heart of the selection.

Redundancy: N+1, 2N, and concurrent maintainability

Cooling redundancy uses the same N notation as the power chain, and it shapes the system as much as capacity does. N is exactly the capacity the design load needs with nothing to spare. N+1 adds one more unit than the load needs, so any single unit can fail or be pulled for service and the load stays cooled. 2N is two complete independent systems, each able to carry the full load alone. The level climbs with the tier the facility is chasing.

Concurrent maintainability is the idea that actually drives the design. It means you can take any single component, a chiller, a pump, a CRAH, a section of pipe, out of service without dropping below the cooling the load needs. That is harder for cooling than for power, because the loop is physical: isolating one pump or one pipe section can strand capacity if the valving and the loop topology were not laid out for it. A plant that is N+1 on the nameplate can still fail concurrent maintainability if it cannot isolate the redundant unit.

The failure that hides in cooling redundancy is the thermal ride-through gap. Power transfers in milliseconds, but a chilled water loop has thermal mass and a chiller takes minutes to restart after a utility blip, so the hall heats up during the gap unless there is thermal storage or enough loop volume to ride through. Redundancy you have not tested under load is redundancy you do not have. The integrated test that proves it is covered by topic in the cooling and airflow guide.

Hybrid air and liquid in the same hall

The AI reality is that most high-density halls are hybrid: direct-to-chip liquid for the hot chips and air for everything else, running side by side in the same room. The cold plate takes the CPUs and GPUs, but the memory, drives, network cards, and power supplies are not on the loop and still need air, commonly the 20 to 30 percent of the rack heat the liquid does not carry. So a liquid hall is not an air-free hall.

That changes how you size and commission the room. You plan two cooling paths at once: a liquid path sized for the chip load and an air path sized for the residual. On a 100 kW rack, the residual can be 20 to 30 kW of air cooling per rack, which is not a rounding error, and teams that budget the CDU and chilled water for the chips and forget the residual air end up with memory throttling in a rack they thought was solved.

The hybrid also shows up across a campus, not just within a rack. A single facility often runs legacy air-cooled halls, mid-density contained rows, and new liquid-cooled AI halls at the same time, all hanging off shared plant. Designing for the mix, rather than assuming the whole site is one type, is the realistic posture for any build going forward. The hybrid rack detail is covered in the direct-to-chip guide.

How do you choose a data center cooling system?

Choosing a cooling system is a matrix of a handful of inputs, and density leads it. Start with the per-rack power the hall will actually carry, because that sets which families are even possible. Then layer on the efficiency target, the water situation, the climate, whether it is a retrofit or a new build, and the budget and tier. The answer is the system that satisfies the density first and then optimizes the rest.

Work the inputs in order. Density picks the family, from room air up to liquid. The PUE target pushes you toward containment, warm supply, and free cooling. The water situation and WUE target push the heat rejection toward evaporative or dry. The climate sets how many free-cooling and dry-cooler hours you get. The retrofit-versus-new question constrains what you can physically install. The budget and tier set the redundancy and the plant type. No single input decides it; the matrix does.

Two practical notes. First, design for the next generation of density, not just the current one, because retrofitting liquid into an air hall is the hard path. Second, whatever you pick, the system has to be commissioned and proven under load before it carries compute; a plant that was never balanced and never failover-tested is a plant you cannot trust. The commissioning sequence is covered by topic in the cooling and airflow guide.

Retrofit, new build, and the climate factor

Adding liquid to an existing air-cooled hall is harder than building it in, and the deciding constraint is usually facility water, not the racks. A new build runs the facility water loop, sizes the plant, and lays out the rows around liquid from the start. A retrofit has to find capacity and a path for facility water in a building that was never plumbed for it, plus floor loading for heavier liquid gear, and that is often the real limit.

Where a retrofit has no facility water to reach, a liquid-to-air CDU is the common bridge: it rejects the chip heat into the room air and leans on the existing air handling, getting liquid onto the chips without a new water plant. The honest caveat is that it relocates the building heat into the room rather than removing it, so the air side has to absorb what the CDU dumps back. Rear-door exchangers are the other low-disruption retrofit, adding liquid capacity at the rack without touching the servers.

Climate sits underneath both cases. A cold or dry climate opens more hours of free cooling and makes dry coolers and waterside economizers carry the load, which can shift the whole design toward water-free rejection. A hot or humid climate leans harder on the chiller and may force evaporative cooling for the peak, which raises the water question. Pick the rejection strategy against the climate the building actually sits in, not a generic assumption.

The AI density jump pulling everything toward liquid

The direction of travel is set by the chips, and it is one way. Each GPU generation lands hotter than the last, the rack power climbs with it, and a larger share of every new build's heat goes to liquid because air ran out of room. The gigawatt-scale AI campuses being built now are designed liquid-first, because at those densities there was never an air option to design around.

This is the trend worth understanding even if your current hall is air. Liquid is becoming the default for serious AI and HPC, the chip vendors publish reference cooling designs because the cooling and the silicon are now sold as one system, and the supply chain and standards are reorganizing around it. Warm-water liquid designs are also pushing heat rejection toward dry coolers and free cooling, because a loop that can run warm can reject heat without a chiller for much of the year.

None of that makes air obsolete. The residual load keeps air in every direct-to-chip rack, and plenty of workloads will never need liquid. The shift is specific: the high-density end has moved to liquid and is not moving back, and the design question for new halls is no longer whether liquid is coming but how much of the load it carries on day one.

What to document

A cooling architecture that was chosen but never written down leaves the operations team guessing at what they inherited. The record is what tells the next engineer which system serves which hall, what density it was sized for, and where the limits are. Capture the system type per hall, where each one captures and rejects heat, the design density and efficiency targets, the redundancy scheme, and the basis for the selection so the reasoning survives the handover.

The table below is the selection-hub view, the one-line summary of each family that belongs in the basis-of-design. Below that, each system type carries its own deeper documentation: the airflow and TAB records for the air side, and the loop, coolant, and leak records for the liquid side, both covered in the sibling guides.

Common mistakes

• Pushing room air cooling past its density, so the cold supply never reaches the rack inlet.
• Skipping containment, then dropping supply temperature to chase hot spots and wrecking efficiency.
• Picking the wrong family for the density, like air-only for an AI or high-density load.
• Optimizing PUE with an evaporative tower while ignoring the WUE and the site's water reality.
• Designing N+1 on the nameplate without a loop that can isolate the redundant unit under load.
• Forgetting the residual air load in a direct-to-chip hall and undersizing the air handling.
• Building for today's rack density with no path to add liquid for the next generation.
• Treating the selection as done at the plant, then never balancing or failover-testing the system.

Field checklist

Standards and references

The central reference for the thermal side is ASHRAE Technical Committee 9.9 and its Thermal Guidelines for Data Processing Environments, which set the recommended and allowable air envelopes, the A1 through A4 equipment classes, and, in the current edition, the liquid-cooling water temperature classes (commonly W1 through W4) that describe the supply temperatures a loop can run. The warmer W classes are what let a loop reject heat with a cooling tower or dry cooler alone, which is why they matter to the rejection choice.

The energy side is ASHRAE Standard 90.4, the energy standard for data centers, which bounds the mechanical and electrical overhead by climate zone, with ASHRAE 90.1 and 62.1 covering the building outside the data hall. PUE and WUE are Green Grid metrics, now widely standardized, used as the operating scorecards for energy and water. Where a facility is chasing a tier, the Uptime Institute Tier standards drive the witnessed redundancy and integrated-test demonstrations, and TIA-942 is the broader telecommunications infrastructure standard for data centers, including environmental and redundancy provisions. The Open Compute Project publishes liquid-cooling reference designs much of the industry aligns to.

Edition numbers and the specific envelope and class values move between cycles, and the density bands, PUE, and WUE figures in this guide are typical ranges, not fixed limits. Confirm the current edition and the actual numbers against the published guideline, the IT equipment manufacturer, and the chip vendor's reference design before citing them on a submittal. ASHRAE TC 9.9 gives the framework; the equipment rating and the project documents control the limit you design and commission to.

Units, terms, and acronyms

Data center cooling carries vocabulary from HVAC, from the IT side, and from commissioning, and the same idea reads differently across a chiller submittal, a Tier document, and a vendor cut sheet. The terms below travel across the whole selection.

CRAC

Computer room air conditioner, a direct-expansion unit with its own compressor and refrigerant circuit

CRAH

Computer room air handler, a chilled water coil unit with no compressor, fed from a central plant

DX

Direct expansion, refrigerant cooling made at the unit and rejected to a condenser outside

In-row / close-coupled

Cooling units placed between the racks to shorten the air path for higher density

Rear-door heat exchanger

An air-to-liquid coil that replaces the rack's back door and cools the exhaust at the rack

Direct-to-chip / cold plate

A cold plate on the CPU or GPU with coolant flowing through it to take heat at the source

Immersion

The whole server submerged in dielectric fluid, single-phase or two-phase

CDU

Coolant distribution unit, which isolates the facility water loop from the technical loop at a liquid-cooled rack

PUE

Power usage effectiveness, total facility energy divided by IT energy; lower is better, 1.0 is the floor

WUE

Water usage effectiveness, liters of water per kilowatt-hour of IT energy; lower is better

Economizer / free cooling

Using cold outside air or water to cool when conditions allow, idling the compressors

kW / ton

Kilowatt of heat load; one ton of cooling is 12,000 BTU per hour, about 3.5 kW

Related tools

Calculators and readiness checks for this work

Compare your options

FAQ

People also ask

• What is bypass airflow in a data center?
• What is direct-to-chip liquid cooling?
• What is the ASHRAE recommended temperature for a data center?
• What is delta-T in a data center?
• What is a CDU in a data center?