Make thermal architecture an early design input
For an air-cooled AI server, thermal design should influence the mechanical architecture before component placement is frozen.
The position of processors, memory, accelerators, storage and power conversion determines whether airflow can be directed cleanly or must fight repeated turns and restrictions. ASHRAE notes that equipment specifications, heat sinks, fan selection and airflow management are primary drivers of server thermal design; facility-level cooling cannot repair a poorly defined chassis path.
A useful early deliverable is a simple heat map: component power, allowable temperature, physical location and the intended cooling route. It exposes conflicts while there is still room to move parts or reserve fan volume. If a platform uses liquid cooling for some devices, define the remaining air-cooled loads separately rather than applying one fan target to the whole system.
As a first air-side estimate, use Q = P / (rho x cp x DeltaT). For an illustrative 10 kW air-cooled load, air density of 1.2 kg/m3, specific heat of 1,005 J/(kg x K), and a 10 K air temperature rise, the result is about 0.83 m3/s, or approximately 1,760 CFM. This is a heat-balance estimate, not a fan selection: recirculation, altitude, heat-source limits, pressure drop, fan efficiency and the actual rack operating point still have to be validated.
- List heat sources by normal and peak power.
- Document temperature limits and measurement locations.
- Reserve space for fans, baffles, connectors and service access.
Build a pressure budget
Static pressure is the force available to push air through resistance. It is consumed by every dense heat sink, fin stack, grille, filter, cable obstruction and abrupt change in cross-section. Pressure loss is also affected by the actual airflow, so a restriction cannot be evaluated independently from the target operating point.
Build a pressure budget for the intended airflow rather than assuming a fan will overcome restrictions. The budget can begin with estimates and be refined with measurement, but it must include the final mechanical features that will be present in production. Record the assumed flow, pressure-drop estimate, tolerance and measurement method for each major element.
- Identify the largest pressure-drop elements and their expected operating flow.
- Avoid unnecessary sharp turns and sudden contractions.
- Check whether parallel branches receive balanced airflow.
- Recalculate the budget when heat sinks, filters, baffles or cable paths change.
Place fans for useful flow
Fans should create a controlled pressure difference across the components they cool. In some enclosures, a push arrangement is practical; in others, pulling air through a dense section gives better distribution. The correct choice depends on the path, the fan curve and the component layout, not a universal rule.
Avoid placing a fan where nearby openings let it short-circuit its own airflow. Use baffles, seals and blanking panels to ensure that the pressure produced by the fan is applied across the intended restriction. In a multi-zone chassis, check that one fan group is not stealing flow from a neighboring zone.
- Keep fan inlets clear of nearby obstructions.
- Use shrouds, baffles and blanking panels where they improve distribution.
- Check bypass leakage around trays, gaps and service openings.
- Review airflow direction with fan replacement and cable access in mind.
Use fan control as part of the thermal strategy
Fixed full-speed operation may provide margin, but it can add unnecessary power, noise and wear under light load. Closed-loop control can match airflow to measured thermal demand, provided the control logic has a defined safe state, adequate response time and enough hysteresis to avoid hunting.
Define sensor placement carefully. A sensor far from the critical component may react too slowly, while a sensor directly in a local hot spot can create excessive speed oscillation without appropriate control tuning. ASHRAE describes fan speed control and cooling zones as ways to match airflow to the components most coupled with each fan. NVIDIA's fan-control documentation also illustrates ramping, hysteresis and temperature-based profiles; these behaviors should be validated with the actual fan and chassis, not assumed from the controller interface alone.
- Set a defined safe fan-speed strategy for missing or invalid sensor data.
- Validate control response and hysteresis during workload transitions.
- Confirm that lower normal speeds still maintain required component temperatures.
- Verify PWM, TACH and alarm behavior at the motherboard or BMC interface.
- Check that cooling zones do not create local bypass or starve a neighboring zone.
Validate the full operating envelope
Thermal validation should include more than one room temperature and one benchmark workload. Inlet conditions, fan failures, dust accumulation, different component options and degraded supply voltage can all change the cooling result.
The objective is not to produce the lowest temperature in a laboratory. It is to show repeatable thermal behavior across the operating conditions that the equipment will actually see. Record the duty point, inlet temperature, fan speed, input power, component temperatures, acoustic result and test configuration so a later mechanical or firmware change can be compared against the same baseline.
Use platform data when it exists. The NVIDIA DGX A100 user guide lists 840 CFM at 80% fan PWM, 22,179 BTU/hr heat output and a 5–30°C operating temperature range for that specific system. These values are not universal AI-server targets; they show why a fan specification should start from the selected server's own thermal data rather than a generic CFM-per-kilowatt assumption. NVIDIA's GPU debug guidance recommends collecting GPU and memory temperature, node fan speed, inlet temperature and power data during troubleshooting and validation. Turn those signals into pass/fail thresholds and time-series records so thermal drift, fan degradation or airflow blockage can be found before throttling occurs.
- Test peak load and realistic inlet temperature conditions.
- Include fan-fault or reduced-airflow scenarios where the product requires resilience.
- Measure rack or chassis inlet temperature, not only room temperature.
- Capture GPU, memory, inlet, fan-speed, power and acoustic observations in one report.
- Repeat the test after filter loading or representative resistance changes.
Turn the thermal design into a tested custom fan
The best thermal architecture still needs a fan configuration that fits the real chassis, control system and production plan.
MEGA Tech can work with AI server teams from the early thermal review. We translate heat maps, pressure budgets, noise limits, available space, voltage, speed range, PWM, FG or RD feedback, redundancy and service requirements into a practical fan specification. The aim is to protect the component temperature limit at the real operating point while avoiding unnecessary fan power and acoustic output.
The configuration can be customized around frame size and thickness, airflow and static pressure, speed, motor and control response, connector and cable length, bearing choice, rotation direction, mounting and environmental protection. This gives established server platforms a reliable, cost-conscious alternative to expensive imported fan assemblies and gives early-stage hardware teams a path to a small first build before production.
We build and test the sample in a representative chassis with the actual heat sinks, baffles, filters, cabling, blanking panels and control logic. Validation covers normal and peak load, loaded airflow path, fan-fault behavior, temperature rise, noise, power, feedback and continuous operation. After the customer approves the configuration, the measured performance window becomes the baseline for stable, repeatable production.
- Define the heat map, pressure budget, inlet condition, noise limit and fault target.
- Customize fan size, pressure, speed, control, feedback, wiring, bearing and mounting details.
- Validate the sample with representative hardware and degraded airflow conditions.
- Carry the approved test baseline into reliable production and service supply.
Thermal design review checklist
A disciplined review connects the thermal target to the physical enclosure, the control logic, the evidence package and the production plan. If a standard fan does not fit the electrical, mechanical or control constraints, discuss that before the prototype cycle, not after a failed validation.
- Heat map, cooling boundary, inlet conditions and component limits are approved by the design team.
- Pressure budget includes final restrictions, tolerances and likely service conditions.
- Airflow path avoids uncontrolled bypass, recirculation and zone imbalance.
- Fan control, sensor placement, fault behavior and degraded-state limits are documented.
- The prototype test plan represents real hardware, workload and inlet temperature.
- The selected fan has a defined sample, measurement baseline and production-control plan.
Build a Time-Series Acceptance Record
A single peak temperature can show that a test failed, but a synchronized time series often explains why it failed.
Record the thermal and airflow signals against the same time base: inlet temperature, GPU temperature, GPU memory temperature, CPU temperature, fan command, actual fan speed, input power and workload state. Add pressure or flow measurements when the test setup permits. The relationships between these signals help separate a blocked path, a slow control loop, a failed fan, an inlet-temperature excursion and a true heat-sink limitation.
Define pass/fail limits before the test. Include steady-state temperature, transient recovery time, maximum allowable fan speed, acoustic limit, fault response, supply-voltage condition and the required degraded state. A test that only reports the lowest temperature at 100% PWM can hide a control problem or an unacceptable service condition.
Repeat the same record after changing the fan, baffle, filter, firmware or heat-sink layout. Keep the test hardware, sensor locations, workload and inlet condition traceable so the engineering team can compare changes without mixing different baselines.
- Synchronize component temperatures, inlet temperature, fan command, actual RPM, power and workload.
- Set acceptance thresholds before running the validation.
- Capture transient response, not only steady-state values.
- Keep sensor locations, hardware, firmware and airflow restrictions traceable across revisions.
Engineer FAQ: AI Server Thermal Design and Fan Validation
Use these questions to turn a thermal concept into a repeatable engineering acceptance plan.
Q: Is room temperature enough to define the test? A: No. Record the server or rack inlet temperature at the equipment, because recirculation and containment can make it different from room average.
Q: Which telemetry should be captured with the fan data? A: At minimum, GPU and memory temperature, CPU temperature where relevant, inlet temperature, fan command, actual fan speed, power and workload state. Add pressure, flow and noise when they affect the design decision.
Q: How do I distinguish a weak fan from a control problem? A: Compare actual RPM and operating-point pressure/flow with the command. A fan that reaches command but still leaves a hotspot points toward distribution, heat-sink or pressure-path issues; a fan that cannot reach command points toward interface, power, fault or mechanical loading issues.
Q: Should the fan curve react directly to the hottest sensor? A: Not automatically. The sensor location, thermal lag, hysteresis, ramp rate and neighboring cooling zones must be tuned together to avoid hunting while still protecting the critical component.
Q: What is a useful degraded-airflow test? A: Define the real failure or blockage case: one fan unavailable, reduced supply voltage, loaded filter, blocked inlet or control-signal loss. Measure whether the remaining system stays within its thermal and acoustic limits.
Q: When is a custom fan sample useful? A: When the chassis needs a particular pressure curve, frame, noise limit, PWM/FG/RD behavior, connector, redundancy response or small first build that standard parts cannot satisfy together.
- Use equipment inlet temperature, not room average only.
- Compare fan command with actual RPM and thermal response.
- Tune hysteresis and ramp behavior with sensor placement and thermal lag.
- Validate defined degraded states before production release.
