Cooling
Room cooling works by filling the room with cold air and hoping it reaches every inlet. At a few kilowatts per rack it does; at twenty or more it does not. The answers move the cooling closer to the heat: into the row, into the rack, and finally onto the processor itself with liquid.
Because room cooling relies on mixing, and dense racks overwhelm it. Schneider Electric's White Paper 130 finds traditional room cooling effective at an average of 1-2 kW per rack, and states that with peak densities of 20 kW per rack or more, traditional cooling without containment, dependent on air mixing, no longer functions effectively. ENERGY STAR makes the same point for room systems designed for 1-5 kW per rack: they cannot keep up, and hot spots develop.
The first steps are containment and blanking, covered in hot and cold aisle containment. The next is to shorten the air path.
Cooling units placed in the row, between the racks, each dedicated to the row it sits in. ENERGY STAR describes row-oriented cooling as each unit cooling one row, mounted among the racks, overhead or under the floor. Schneider Electric lists the benefits: shorter airflow paths that cut fan power, the full rated capacity usable, and capacity and redundancy, N+1 or 2N, targeted at the rows that need them.
That targeting is useful in mixed rooms. ENERGY STAR gives the example of putting mission-critical servers in a row with redundant cooling while less important loads sit in a row without it. For existing data centres, Schneider Electric suggests considering row-based cooling when deploying loads of 5 kW per rack and above.
Cooling built into, or attached to, the rack itself, so the airflow path is as short as it gets. Schneider Electric states that rack-based paths are shorter and exactly defined, immune to installation variation and room constraints, and achieve the highest density of the three architectures, up to 50 kW per rack. ASHRAE's liquid cooling paper mentions the closely coupled option for existing raised-floor rooms: supplementing or replacing raised-floor cooling with a coil such as a rear door heat exchanger.
When the heat has to be taken from the chip rather than from the air. ASHRAE TC 9.9's paper on the emergence of liquid cooling describes direct on-chip cold plates, which most of today's designs can run on W32 facility water, and immersion, in which the equipment sits in a fluid. It also states that there is a rack density, different for each solution, at which the first cost of liquid cooling falls below that of air-cooling the same load.
More power in fewer racks, and more heat from the UPS that feeds them. Schneider Electric's cooling-load worksheet counts UPS and power distribution losses as part of the room's heat; at high density the UPS is usually larger and its losses larger with it. Liquid systems also bring pumps and coolant distribution units that need power, and a cooling path that must keep running through a utility failure, which belongs in the UPS and generator plan.
Distribution changes too: high-density racks draw more per circuit and usually per rack PDU; see rack PDU types. The heat arithmetic is in the server room cooling sizing guide.
Predictability and redundancy become the hard part. Schneider Electric notes that in uncontained rooms the assurance of cooling unit redundancy becomes a very complicated analysis that is difficult to validate, while row-based designs let N+1 or 2N redundancy be targeted at specific rows. Shorter air paths also mean the full rated capacity of each unit can be used, rather than part of it being lost to air that bypasses the equipment.
The power path changes in step. Dense racks are usually fed by higher-capacity circuits and rack PDUs, and the UPS that feeds them carries more load in fewer places, so a single cooling fault near those racks affects a larger share of the UPS load. That is a reason to give the densest rows the redundant cooling, as ENERGY STAR's example of a mission-critical row suggests.
By the density you have now, the density you expect, and the room you are in. A room averaging a few kilowatts per rack with a few dense racks can often take in-row units next to those racks. A new high-density deployment may start with rack-based or rear-door cooling and leave space for liquid. Processor-level heat that air cannot carry calls for cold plates or immersion, with the facility water the paper describes.
Cooling design, installation and maintenance are delivered through specialist cooling subcontractors. On the power side, UPS capacity planning and service for data centres covers the UPS that feeds the racks.
There is no single line. Schneider Electric finds room cooling effective at an average of 1-2 kW per rack and ineffective without containment at 20 kW per rack or more.
Cooling units placed within the row of racks, each dedicated to that row, so cold air travels a short distance to the equipment inlets.
Schneider Electric gives up to 50 kW per rack for rack-based cooling.
Not always. ASHRAE states that above a certain rack density, which varies by solution, liquid cooling has a lower first cost than air-cooling the same load.
Its pumps and coolant distribution units need power, and cooling must continue through a utility failure, so they belong in the UPS and generator plan.
Sources, checked 2026-10-04:
Tell us the site, the UPS and the timing. We reply and quote within 4 business hours, or call 1 (438) 800-0382
Brand names are trademarks of their owners and are used only to identify the equipment we service. We are not the manufacturer and are not affiliated with it.