CARDINALPlan GPU power and cooling around the complete server and its installation. A card's wattage does not describe wall power, rack capacity or cooling compatibility. Confirm the exact server configuration with its manufacturer and your facility operator before buying accelerators.
Start with the measurement boundary
Power figures answer different questions. The accelerator rating helps describe one component. A server maximum concerns the assembled system. A metered average describes a particular workload over time. A rack allocation describes what the facility agrees to support.
Put those labels beside every figure in your purchasing worksheet. Otherwise, a seller's GPU rating can become an unsupported assumption about the complete installation. Ask whether each number is a published maximum, a configured limit, an estimate or a measurement.
Also record where the measurement was taken. Device telemetry and a server's AC input do not share the same boundary. If you compare offers, compare equivalent boundaries and operating conditions. Keep units visible: watts measure power, while kilowatt-hours measure energy.
The NVIDIA DGX H100/H200 guide illustrates this distinction. It lists a maximum system power of 10.2 kW and specifies the input requirements for its power supplies. That is a documented complete-system reference, not a rating for every server containing H100 or H200 accelerators. Source: DGX H100/H200 system power specifications
Passive cooling still needs controlled airflow
A passive GPU heatsink depends on airflow supplied by the system. “Passive” does not mean that the card can operate without forced cooling. The enclosure and its supported cooling arrangement are part of the intended installation.
NVIDIA lists the L40S as a passive, dual-slot accelerator with a maximum power consumption of 350 W. These facts identify a cooling requirement; they do not establish compatibility with an arbitrary desktop enclosure. Source: NVIDIA L40S full specifications
Ask the server supplier to identify the supported GPU positions and the necessary fan, shroud and cable configuration. Ask for the applicable service-manual section. A card fitting physically into a slot is only one part of that review.
Thermal limits can depend on the whole configuration. Dell's R760xa thermal matrix distinguishes processor selections and cooling requirements. It also requires six system fans across the documented configurations. A statement that a server family supports GPUs is therefore less useful than confirmation of the offered build. Source: Dell R760xa thermal restriction matrix
For used equipment, request photographs of the complete internal arrangement and an itemized list of cooling parts. A removed shroud or missing fan can be easy to overlook in an offer centered on the accelerators. Ask the integrator to resolve undocumented modifications before installation.
Module cooling is a system decision
An SXM or OAM module belongs to a compatible platform. Do not treat it as a PCIe card with a different bracket. Your request should identify the baseboard, server and cooling assembly, or state that you need those supplied together.
If an offer involves liquid cooling, ask for the exact server integration requirements and facility interface. Record who supplies the required equipment and who commissions it. A cooling label alone does not describe a ready-to-install system.
This guide is a purchasing worksheet, not an installation procedure. Qualified installation staff should follow the server manufacturer's instructions. Changes to electrical distribution or cooling infrastructure belong with the responsible facility team.
Build a transparent power budget
Separate component arithmetic from the final facility requirement. The following example is hypothetical. It shows how to expose assumptions, not how to certify a particular server or select its power supplies.
Assume four accelerators are each budgeted at 350 W. Assume another 600 W of DC demand for the rest of the server. If power conversion is 94% efficient at that operating point, estimated AC input is 2,000 ÷ 0.94, or approximately 2,128 W.
The 600 W allowance and 94% efficiency are illustrative assumptions. Replace them with configuration-specific evidence. Do not add conversion losses again if your starting figure already measures complete-server AC input.
Scroll across the diagram to read every label
Download diagram| Budget item | Calculation | Illustrative result |
|---|---|---|
| Accelerators | 4 × 350 W | 1,400 W DC |
| Other server components | Assumption requiring verification | 600 W DC |
| Total component demand | 1,400 + 600 | 2,000 W DC |
| Estimated input | 2,000 ÷ 0.94 | Approximately 2,128 W AC |
| Conversion losses | 2,128 − 2,000 | Approximately 128 W |
The diagram and table use rounded values. They exclude facility cooling equipment and other rack devices. They also do not model short transients, power-factor effects or behavior after a component failure. Ask the manufacturer and facility operator which additional requirements apply.
Separate energy cost from capacity planning
For an energy estimate, multiply average input in kilowatts by operating hours. If the hypothetical server averaged 2.128 kW for 200 hours, it would use approximately 425.6 kWh. This is arithmetic using an assumed average, not a workload forecast.
To estimate the electricity component of operating cost, multiply those kilowatt-hours by the applicable energy tariff. Keep the tariff date and currency visible. Facility charges may use a different billing structure, so check the actual hosting proposal.
A low average does not establish sufficient installed capacity. Procurement should retain separate fields for expected average consumption, supported maximum demand and facility allocation. That separation makes later measurements useful without weakening the installation requirement.
Ask what happens when a component fails
Do not calculate usable redundant capacity by simply adding every power-supply label. Ask which failure the configuration supports and what performance remains afterward. The answer should refer to the exact server and power arrangement.
Dell documents 5+1 power-supply redundancy for the XE9680, meaning one power supply is redundant. Its guide also describes configuration-dependent power-brake behavior. This is a concrete reason to distinguish normal operation from a reduced-power condition. Source: Dell XE9680 power-supply specifications
Put the question in operational terms: if one supported power component fails, can the agreed workload continue at its required service level? If not, record the expected limitation. Have the facility operator review power-feed assumptions separately from server-level redundancy.
Ask whether the seller's performance evidence used the same power limits and supply configuration as the offer. A benchmark from a different configuration cannot establish the performance you will receive. Keep any mismatch visible until the supplier resolves it.
Confirm the installation before committing
Send the facility operator the server's complete specification and the proposed quantity. Include physical dimensions, loaded weight, mounting requirements and connection details. Ask for written confirmation of the supported installation rather than a general statement about available rack space.
Record who checks the rack, receives the shipment and performs installation. Confirm whether the offer includes the correct rails and required accessories. A missing low-cost component can still delay a high-value system.
Cooling review should identify the supported inlet conditions and the proposed rack arrangement. Use the applicable OEM requirements rather than a generic temperature target. Do not extrapolate one model's allowed environment across a mixed rack.
For a staged deployment, document the initial configuration and the intended expansion separately. The first server fitting within the allocation does not prove that the final populated rack will fit. Ask which approvals must be repeated when the configuration changes.
Request the complete supported configuration
A useful offer names the accelerator variant, exact server build, included cooling components and applicable documentation. It also names the party responsible for integration. Avoid leaving those responsibilities spread across several unanswered email threads.
Use the GPU request-for-quote template to collect the details. Add the existing server service identifier where appropriate, but do not publish private asset records. Ask the supplier to mark unknown or excluded items explicitly.
When comparing offers, separate the accelerator price from the cost of reaching the required operating configuration. The offer comparison worksheet provides blank fields for those costs. It contains no market-price assumptions.
The worksheet calculates a subtotal after you enter quantity and unit price. Its total remains blank until every cost field contains a number. Enter zero only when you have confirmed that a cost does not apply or is included elsewhere.
Cardinal sources hardware and checks the documents sellers provide. The relevant manufacturer, integrator and facility operator must confirm installation suitability. Include unresolved power or cooling questions in your quote request so they remain part of the sourcing brief.
Common questions
Can a passive datacenter GPU run in a workstation?
Only if the complete installation is supported for that accelerator and its cooling requirements. A compatible-looking slot does not establish sufficient airflow, power delivery or manufacturer support. Request confirmation for the exact enclosure and configuration.
Is GPU TDP the server's wall power?
No. A GPU rating concerns the accelerator. Complete-server input includes other components and power conversion. Use the server manufacturer's input specification or a clearly identified measurement when discussing facility requirements.
Does a larger power supply guarantee compatibility?
No. Capacity alone does not establish the supported connections, cooling arrangement or failure behavior. Request the OEM configuration requirements and confirm the installed parts. Treat improvised changes as unresolved until the responsible integrator reviews them.
Should I buy cards or a complete GPU server?
Buy individual cards when you have a verified compatible host and a clear integration plan. Request a complete supported system when those pieces remain open. Compare both paths at the same operating boundary, including missing accessories and integration work.