Start with the cooling problem, not the fan
AI server fan selection is a system problem: concentrated heat, restrictive flow paths and limited thermal margin must be solved together.
AI servers make fan selection difficult because their heat sources are concentrated and their airflow paths are restrictive. A free-air CFM value can be useful for comparison, but it does not describe how a fan will behave behind a processor heat sink, accelerator module, drive cage, filter or cable bundle.
First identify the cooling boundary. In an air-cooled server, the fan must move air through the chassis and across the defined heat sources. In a liquid-cooled or hybrid platform, the fan may mainly serve power supplies, storage, networking, heat exchangers or residual air-cooled components. The fan specification should follow that boundary rather than assume every AI server has the same air-cooling duty.
Start by documenting the heat to be removed, the allowable component temperatures, the expected inlet-air temperature and the physical airflow route. NVIDIA guidance for GPU-ready facilities emphasizes that higher-power GPU servers need more airflow and that gaps and recirculation must be controlled; those facility requirements still need to be translated into the server's own fan and chassis operating point.
- Identify the heat-generating components and their worst operating condition.
- Map the intended inlet, outlet and every major restriction between them.
- Record available fan dimensions, connector, voltage and control constraints.
- Separate server-level airflow from rack- and facility-level cooling targets.
Translate heat load into an airflow target
For an air-cooled enclosure, the required airflow depends on the heat load and the permitted rise in air temperature. A first-order estimate can use Q = P/(rho x cp x ΔT), but this is only a starting point: airflow must still reach the hot components and overcome the resistance of the chassis.
Use the estimate to define a target range rather than a single CFM number. The real operating point will move as filters load, fan speed changes and optional hardware is installed. Do not copy a facility-level airflow-per-kilowatt figure directly into a server fan specification; the server must be evaluated with its own pressure path and internal heat sources.
A useful reference point is the Open Compute Project North Dome specification: it gives an example system target of 0.13 CFM/W, a desired system ΔT greater than 13.9°C, testing up to a 35°C inlet condition and N+1 fan redundancy. These are platform-specific design values, not universal AI-server requirements, but they show the level of detail a fan specification should carry before a sample is built. NVIDIA's current DGX SuperPOD planning guidance also uses a facility-level airflow figure for a particular rack architecture; that figure should not be substituted for the server's own fan duty.
- Define normal and worst-case ambient conditions.
- Allow margin for distribution losses and bypass leakage.
- Check whether each local hot spot receives enough flow, not only whether the chassis total is adequate.
- Label every airflow-per-power figure by platform, cooling layer and test condition.
Use the fan curve and system curve together
A fan curve shows the airflow a fan can provide at different static pressures. A system curve describes how much pressure the enclosure demands as airflow increases. Their intersection is the likely operating point.
Dense GPU and accelerator platforms commonly shift that intersection toward higher pressure. A fan chosen only for maximum free-air CFM can therefore deliver less useful airflow than a lower-CFM model with a stronger pressure curve.
- Request or measure P-Q curves at the intended voltage and control setting.
- Estimate the pressure contribution from heat sinks, guards, filters and narrow passages.
- Recheck the operating point when the mechanical layout changes.
Select the electrical and control interface
Voltage, current, start-up behavior and control interface must be treated as enclosure requirements, not catalogue afterthoughts. The fan supply must tolerate the measured or specified transient load at start-up, and the controller must support the chosen speed-control method without losing fault visibility.
For server applications, tachometer feedback and PWM control are often valuable because they allow the platform to verify fan health and tune noise and power against thermal demand. Whether a custom lead length, connector or signal format is required should be decided early, before validation samples are built. The control range also needs a lower-speed thermal test; a fan that responds correctly to PWM is not automatically safe at every commanded duty cycle.
- Confirm rated voltage and the permitted operating range.
- Use measured or supplier-specified start-up current when sizing redundant fan banks.
- Define required feedback, alarm and speed-control signals with the electronics team.
- Test the minimum, nominal and maximum commanded speeds in the representative chassis.
Balance acoustics, redundancy and service
A fan selection that passes a thermal test at full speed may still be unsuitable if the acoustic level, vibration or power use is unacceptable in normal operation. Fan speed control can reduce these penalties, but only if the thermal margin is verified at lower speeds.
Redundancy is a system decision, not simply a fan count. Multiple fans can maintain cooling after one failure only if the remaining fans, control logic, power supply and airflow path have been validated in that degraded state. Tool-free access, clear airflow direction and replacement procedures are equally important in 24/7 equipment.
- Test both normal and defined degraded/failure fan conditions.
- Measure component and rack-inlet temperature, not only outlet-air temperature.
- Confirm power and control behavior when one fan or one control channel is unavailable.
- Review fan replacement access before releasing the mechanical design.
Validate with a representative enclosure
Bench tests with an open fan do not reproduce a server. Validate samples in a mechanically representative chassis with the final heat sinks, cabling, blanking panels and software control logic whenever possible.
A useful validation report combines temperature data, fan speed, input power, acoustic results and observations about leakage or recirculation. This gives engineering teams a clear basis for deciding whether to adjust fan selection, baffle geometry or operating logic.
- Run steady-state and transient thermal tests.
- Check the hottest component under degraded airflow conditions.
- Record the tested configuration so later changes can be compared fairly.
From AI server requirements to a production-ready fan
The final fan should be developed around the server's real chassis, not selected from a free-air number and left for the enclosure team to solve.
MEGA Tech can work with AI hardware teams from the first cooling review. We translate heat load, GPU and power-shelf airflow, pressure loss, available space, noise limit, voltage, speed range, control signal, feedback, redundancy and service requirements into a practical fan specification.
The configuration can be customized around frame size and thickness, airflow and static pressure, speed, PWM, FG or RD feedback, connector and cable length, bearing choice, rotation direction, mounting and environmental protection. This supports established server platforms that need a reliable, cost-conscious alternative as well as new AI products that need a small first build before production.
We build and test a sample in the representative chassis with the actual heat sinks, cabling, blanking panels, filters and control logic. Validation covers operating point, temperature rise, noise, power, speed response, fan-fault behavior and continuous operation. After approval, the same configuration can move into stable volume production with a clear performance baseline and consistent supply.
- Define the AI server heat load, pressure path, electrical interface and reliability target.
- Customize mechanical, electrical, control, feedback, bearing and mounting details.
- Validate the sample with representative hardware, software control and degraded airflow.
- Transfer the approved configuration into reliable, repeatable production.
Engineer FAQ: AI Server Fan Selection
Use these questions to keep platform-specific thermal values from becoming unsupported universal rules.
Q: Does 0.13 CFM/W apply to every AI server? A: No. It is a platform-specific example with stated boundary conditions. CFM/W must be tied to the server architecture, inlet condition, allowed temperature rise, pressure path and whether fan power is included.
Q: What should thermal margin include? A: Define the maximum component or inlet limit, worst workload, inlet temperature, sensor uncertainty, airflow distribution and degraded-fan condition. A margin percentage without those assumptions is not a complete acceptance criterion.
Q: Are peak and average GPU loads interchangeable for fan selection? A: No. Use the workload and time window that create the thermal worst case, then check transient response when power, fan command or workload changes.
Q: Does a liquid-cooled AI server eliminate the need for server fans? A: Not automatically. Power supplies, storage, networking, control electronics and heat exchangers can still reject heat to air and need their own airflow path.
Q: What does N+1 actually require? A: Test the remaining fans, controller, power rail and airflow path after one defined failure at the representative system resistance and thermal load. A fan count alone does not prove thermal redundancy.
Q: When is a custom fan sample useful? A: When the product needs a particular frame, pressure, noise, PWM/FG/RD interface, connector, redundancy behavior or small first build that standard parts cannot satisfy together.
- Platform-specific CFM/W and Delta-T values need their original conditions.
- Thermal margin must include workload, inlet temperature, sensors and degraded operation.
- Peak, transient and steady-state behavior should be tested separately.
- Liquid cooling changes the air-cooling boundary; it does not always remove it.
- N+1 is a validated degraded state, not just a quantity of fans.
AI server fan-selection checklist
Before releasing a fan specification, confirm that the selected model is supported by system evidence rather than a single catalogue value. A cooling partner can help compare standard options and discuss custom mechanical, electrical or control details when the enclosure is constrained.
- Cooling boundary, heat load, ambient condition and temperature limits are documented.
- The fan curve meets the estimated system pressure at the required flow.
- Electrical, connector, feedback and control requirements are compatible.
- Redundancy and degraded-airflow behavior are defined.
- Thermal, acoustic, failure-mode and service tests have a defined plan.
