GPU Data Center
Cooling Requirements
How to plan cooling for high-density GPU racks: separate the four power figures that get confused with each other, understand what the OEM system actually requires, and know what your site engineering has to confirm before a design is real.
Updated
GPU cooling must be designed from the complete server and rack heat load, not from GPU board power alone. The correct system depends on rack density, the OEM thermal requirements for the exact configuration, facility water conditions, redundancy, climate, and the existing heat-rejection plant. High-density deployments often require direct liquid cooling, but the final design must follow the selected server specification and site engineering.
What this guide is. Planning guidance, so you can scope the problem and hold a competent conversation with a mechanical engineer. It is not an interactive planner, it performs no calculation, and it does not replace the OEM system specification, local code, a site survey, or stamped mechanical and electrical engineering. Every real deployment needs all four.
Four Power Figures, Not One
Most cooling sizing errors come from using one of these where another belongs. They increase in that order, and only the last is what the mechanical plant sees.
| Figure | Covers | Where you get it | What it is for |
|---|---|---|---|
| GPU board power | The accelerator alone | Manufacturer specification, and often a configurable range rather than a figure | Comparing accelerators. Not a cooling input on its own. |
| Server input power | Accelerators, CPUs, memory, storage, network adapters, fans, and power-supply losses | The OEM system specification for your exact configuration | The real per-chassis heat load. This is the number the OEM qualifies its thermal solution against. |
| Rack design load | Provisioned capacity per rack, including diversity and headroom | Your own design decision, constrained by the hall | Electrical distribution, containment, and airflow or coolant provisioning per rack. |
| Facility input power | IT load plus cooling and electrical losses across the site | Calculated from IT load and your site's efficiency | Utility service, generation, and heat-rejection plant sizing. |
Why board power is the wrong starting point
An accelerator's published board power excludes everything else in the chassis, and for many parts it is published as a configurable range rather than a single figure, with the design value set by the OEM system. Two servers holding the same eight accelerators can present materially different input power and different airflow or coolant requirements, because the rest of the configuration differs.
Start from the OEM system specification for the configuration you intend to order. Board powers for the accelerators themselves, with the manufacturer page and check date for each, are in the GPU specifications database (18 records), and are useful for comparison but not for sizing.
Turning a load into a facility figure
The power and cooling calculator steps from accelerator count through server, rack, cooling, and facility input power, with its formulas and assumptions shown alongside the result. Use it to produce the numbers you take into a mechanical conversation.
Where the Authoritative Guidance Now Lives
The reference most cooling material still cites has been superseded. This matters if you are working from figures quoted second-hand.
ASHRAE Technical Committee 9.9 is the body responsible for thermal guidance on data centers and IT equipment. TC 9.9 records that since 2024 its guidance has been consolidated into the ASHRAE TC 9.9 Datacom Encyclopedia, which carries the content formerly published as the printed Datacom Series books Thermal Guidelines for Data Processing Environments, Design Considerations for Datacom Equipment Centers, and Liquid Cooling Guidelines for Datacom Equipment Centers.
Two things follow that make older quoted figures unsafe to rely on. TC 9.9 states the liquid-cooling content was almost entirely rewritten and now introduces a new set of S Classes for the supply temperature of the Technology Cooling System coolant. TC 9.9 has also published a Water-Cooled Servers white paper whose stated purpose includes correcting misunderstandings in the last printed edition of the liquid-cooling book.
SLYD does not reproduce ASHRAE's tables on this page. The current edition is a subscription resource and it is the authority. Temperature envelopes, coolant chemistry limits, filtration ratings, and the S Class definitions should be read from the current Encyclopedia by whoever is stamping the design, not taken from a figure quoted on a vendor page against a superseded edition.
Also applicable
- Local electrical and mechanical code, which governs regardless of any guideline
- The OEM system specification, which is contractual for warranty and support purposes
- Water treatment and discharge rules for the jurisdiction, where an open loop or evaporative rejection is involved
- Structural loading limits, which liquid-cooled and immersion deployments can reach before density limits do
What to ask your engineer to confirm
- Which ASHRAE class and which S Class the proposed design targets, from the current edition
- Approach temperatures and the resulting supply and return conditions on both loops
- Coolant flow rate and pressure drop per rack against the OEM requirement
- Redundancy of pumps, coolant distribution units, and heat rejection, and the failure mode of each
- Water chemistry, filtration, treatment, and the monitoring regime that maintains them
The Four Approaches and What Each Demands
Described by what they require of the building and of the hardware, because that is what decides whether one is available to you.
Air cooling
Existing hallsCold supply air through the chassis, hot air returned to the room and on to the cooling units. The density it supports depends on supply air temperature, available airflow per rack, containment quality, and how well recirculation is controlled.
Requires of the building
- Sufficient airflow per rack at the design supply temperature
- Effective hot and cold aisle separation to prevent recirculation
- Floor or overhead distribution sized for the airflow, not just the load
Rear-door heat exchangers
Retrofit pathA liquid-to-air coil in the rack door captures hot exhaust before it enters the room. Passive versions rely on server fans; active versions add their own. Servers themselves stay air-cooled and unmodified.
Requires of the building
- Chilled or condenser water piped to each rack position
- Leak detection and containment at the rack
- A rejection plant that can take the captured load
Direct liquid cooling
Often mandatoryCoolant is pumped through cold plates mounted on the accelerators and other hot components. A coolant distribution unit separates the clean technology cooling loop from the facility water loop and transfers heat between them. Some accelerator platforms are only offered this way, in which case the decision is made for you.
Requires of the building
- Facility water at a temperature that lets the coolant distribution unit hold the required supply condition
- Flow rate and pressure at each rack matching the OEM requirement
- Coolant chemistry, filtration, and treatment maintained to specification, with monitoring
- Redundancy in pumps and distribution units, since loss of flow is a fast failure
- Leak detection, containment, and a defined response
Immersion cooling
SpecialistServers are submerged in a dielectric fluid, either single-phase with the fluid pumped through a heat exchanger, or two-phase where the fluid boils at the component surface. The chassis and the room both change shape completely.
Requires of the building and the hardware
- OEM support for immersion of the specific hardware, including warranty treatment
- Structural capacity for filled tanks, which can be the binding constraint
- Fluid handling, storage, and a serviceability model your staff can actually operate
- For two-phase, containment appropriate to the fluid and attention to its regulatory status
Two Loops, Joined by a CDU
The single most useful thing to understand about a liquid-cooled deployment, and the thing most often collapsed into one system in summaries.
| Aspect | Technology cooling system | Facility water system |
|---|---|---|
| Where it runs | Coolant distribution unit to the cold plates inside the servers | Building side, from the distribution unit to the heat-rejection plant |
| Control tightness | Closely controlled supply temperature and flow, to an OEM requirement | Looser, and driven by climate and plant capability |
| Chemistry and filtration | Strict, because contamination or corrosion reaches the cold plates | Less strict, but still treated and monitored |
| Who owns it in practice | Usually specified against the IT hardware requirement | Usually an existing building system you have to work within |
| Failure consequence | Loss of flow is a fast thermal event for the hardware | Loss of rejection is slower, and buffered by loop volume |
What the CDU actually decides
The coolant distribution unit sets what technology-loop supply temperature you can hold given your facility water temperature, and its capacity and redundancy set how much load a group of racks can carry. Two design questions follow, and both are easy to defer and expensive to get wrong:
- Approach temperature. The distribution unit can only deliver technology-loop supply a certain amount above facility water temperature. Warm facility water may make economization possible and remove chillers, but it also raises the floor on what the servers receive. That interaction decides both the plant design and whether the servers stay in specification.
- Placement and redundancy. In-rack, in-row, and centralized distribution units differ in failure domain, in how much floor they consume, and in how much piping runs through the white space. Decide the failure domain you are willing to accept before choosing the topology.
Working Through the Selection
In order. Most of the answer is determined by steps one and two, before any comparison of cooling technologies.
| Step | What to establish | Why it comes here |
|---|---|---|
| 1. Read the OEM specification | Thermal requirement, cooling method, airflow or coolant flow, inlet conditions, and board-power configuration for the exact system you intend to order | If the platform is liquid-cooled only, the comparison is over. If it has an air-cooled variant, that variant may fit a hall you already have. |
| 2. Survey what the site can actually deliver | Available power, supply air temperature and airflow, containment, facility water temperature and capacity, rejection plant headroom, structural capacity, and floor space | The site eliminates options faster than any preference does, and finding this out late is what forces redesigns. |
| 3. Set the rack design load | Provisioned load per rack, with diversity and headroom, and the number of racks | This is the number that electrical distribution and cooling provisioning are both sized from. |
| 4. Choose the approach | Air, rear-door, direct liquid, or immersion, and the distribution-unit topology if liquid | By this point the choice is usually narrow, and it is driven by steps one and two rather than by the merits of the technologies in the abstract. |
| 5. Fix redundancy and failure modes | Redundancy of pumps, distribution units, and rejection; ride-through on loss of flow; leak detection and response | A liquid-cooled rack has a much shorter thermal ride-through than an air-cooled one, so this is a design input rather than an operational detail. |
| 6. Get it engineered and stamped | Mechanical and electrical design, code compliance, commissioning plan | Nothing above is a design. It is the scoping that lets an engineer produce one efficiently. |
On PUE
Worth understanding, and worth not treating as a property of a cooling technology.
Power usage effectiveness is total facility power divided by IT power. It is an outcome of a whole site: climate, supply and return temperatures, whether economization is available, the age and control of the rejection plant, load factor, and electrical topology all move it, and so does the boundary you choose to measure at.
A PUE figure attached to a cooling technology, with no site behind it, is not something you can design to. Liquid cooling can improve PUE, principally by removing fan power and by allowing warmer facility water that widens the window for economization, but by how much is a property of your site rather than of the technology. SLYD publishes no PUE figures per cooling approach on this page for that reason.
Model your own site. The power and cooling calculator takes PUE as an input you set, and shows how it propagates into facility input power, so you can see the sensitivity rather than adopt a published number.
Frequently Asked Questions
Common questions about cooling high-density GPU deployments
At what rack density do I need liquid cooling?
There is no universal threshold, and treating one as universal is how cooling projects go wrong. The answer comes from the specific server you are deploying and the facility you are deploying it into. Some accelerators are only offered on liquid-cooled platforms, in which case the decision is already made. Others are offered in air-cooled variants precisely so that existing halls can take them. Read the OEM system specification for the configuration you intend to buy, then check it against your available supply air temperature, airflow, containment, and heat-rejection capacity.
Is GPU board power the same as the heat I have to reject?
No, and confusing the two is the most common sizing error. GPU board power covers the accelerator alone. Server input power adds CPUs, memory, storage, network adapters, fans, and power-supply losses. Rack design load is what you provision per rack including diversity and headroom. Facility input power adds the cooling and electrical losses on top of the IT load. Each of these is larger than the one before it, and the number you design the mechanical plant around is the last one.
What is the current ASHRAE guidance for data center cooling?
ASHRAE Technical Committee 9.9 is the relevant body. Since 2024 its guidance has been consolidated into the ASHRAE TC 9.9 Datacom Encyclopedia, which supersedes the printed Datacom Series books that previously carried Thermal Guidelines for Data Processing Environments and Liquid Cooling Guidelines for Datacom Equipment Centers. TC 9.9 states the liquid-cooling content was almost entirely rewritten and introduces a new set of S Classes for Technology Cooling System supply temperature. Because that is a subscription resource and it is the authority, SLYD does not reproduce its tables. Obtain the current edition rather than relying on figures quoted from an older one.
What is the difference between the technology cooling system and the facility water system?
They are two separate loops joined by a coolant distribution unit. The technology cooling system is the clean, closely controlled loop that runs to the cold plates inside the servers, and it is the one with strict requirements on chemistry, filtration, and supply temperature. The facility water system is the building-side loop that carries the heat to the rejection plant, and it operates to looser conditions. The coolant distribution unit transfers heat between them and isolates the server loop from the building loop.
Can this guide size my cooling system?
No. It explains what drives the design and what you need to establish, which is enough to have a competent conversation and to avoid the common errors. It does not replace the OEM system specification, local code, a site survey, or stamped mechanical and electrical engineering. Any real deployment needs all four.
Does liquid cooling improve PUE?
It can, but PUE is an outcome of a whole facility rather than a property of a cooling technology. The same cooling approach produces different PUE depending on climate, supply and return temperatures, whether the site can use economization, the age and control of the rejection plant, and the load factor. A PUE figure quoted against a cooling technology, with no site behind it, is not something you can design to. Use your own site conditions.
Methodology and limits
What this guide is
Planning guidance. It is written to help you scope a cooling problem, separate the figures that get confused with each other, and know what to establish before engaging an engineer. It performs no calculation and is not an interactive planner.
What it does not replace
- The OEM system specification for the configuration you intend to order, which is the governing thermal requirement and is contractual for warranty and support
- Local mechanical, electrical, and plumbing code
- A site survey of what your facility can actually deliver
- Stamped mechanical and electrical engineering
Why specific ASHRAE values are not reproduced here
The previous version of this page published inlet-temperature envelopes, coolant pH, conductivity, chloride, suspended-solids limits, and filtration micron ratings attributed to ASHRAE. Those figures came from the printed Datacom Series books. ASHRAE TC 9.9 records that since 2024 that content has moved into the subscription Datacom Encyclopedia, that the liquid-cooling guidance was almost entirely rewritten, and that it now introduces S Classes for Technology Cooling System supply temperature. TC 9.9 has also published a white paper whose stated purpose includes correcting misunderstandings in the last printed edition.
SLYD cannot read the current Encyclopedia from a public source, and republishing superseded thresholds against a live standard would be worse than omitting them. The guide therefore names the authority precisely and directs the reader to it.
Why no universal density thresholds or PUE figures
A density at which air cooling stops working is a property of a server and a hall, not a number. PUE is an outcome of a whole facility, not a property of a cooling technology. Publishing either as a general figure invites a reader to design against something that does not apply to their site.
Sources and basis
What supports each part of this guide, and when it was last checked.
That ASHRAE TC 9.9's guidance has been consolidated since 2024 into the Datacom Encyclopedia, superseding the printed Datacom Series books; that the liquid-cooling content was almost entirely rewritten; that it introduces S Classes for Technology Cooling System supply temperature; and that a TC 9.9 white paper exists to correct misunderstandings in the last printed edition.
The Encyclopedia itself is a subscription resource and was not read. SLYD therefore reproduces none of its temperature envelopes, chemistry limits, filtration ratings, or class definitions.
ASHRAE TC 9.9 committee site, Datacom Encyclopedia and publications sections
That the current authoritative edition is distributed as an online subscription rather than as a printed book.
Content is gated behind a subscription. Only the existence and access model were verified.
Accelerator board powers, cooling designations, and form factors referenced by this guide.
Board power is the accelerator alone and several parts publish a configurable range. The governing thermal figure is the OEM system specification, not the accelerator page.
The separation of GPU board power, server input power, rack design load, and facility input power, and what each is used for.
A standard distinction in data center design. Reproduced here because collapsing these four figures is the most common sizing error on this subject.
The description of the four cooling approaches, the two-loop model and the role of the coolant distribution unit, and the ordering of the selection process.
SLYD's guidance, framed around what each approach requires of the building and the hardware. Reviewed for accuracy; not a substitute for engineering and not a manufacturer or standards claim.
The absence of universal rack-density thresholds, heat-flux thresholds, per-technology PUE figures, heat-capture percentages, and dielectric fluid prices.
Site-specific outcomes and volatile commercial facts. The previous version published all of them without a source, and a reader could have designed against figures that do not apply to their facility.
Planning a High-Density Deployment?
Tell us the system you are considering and what your site can deliver, and we will work through the constraints with you before the design is committed.