Emulating Processor Power for Thermal Testing: Why Constant Power

A heat sink is not specified in volts. Neither is a cold plate, a vapour chamber or a liquid loop. They are specified by the heat they can remove and the temperature rise they allow while removing it. So when a thermal module is validated, the electrical load standing in for the processor has one job: hold a stated number of watts, steadily, for as long as the test runs.

That sounds trivial until you look at what a resistive thermal simulator does as it warms up.

Why the mode matters more than the supply

A resistive thermal simulator’s resistance rises as it heats, so the mode the supply is operating in decides whether the thermal load stays at its rated value. How much it drifts is set by the heater element’s temperature coefficient of resistance, which is why the element matters as much as the supply: a nickel-chromium element moves relatively little over a working range, while a copper or nickel one moves a great deal. The three cases diverge in direction regardless:

• Constant voltage. As resistance rises, current falls, and power falls with it. The module under test is given a lighter thermal load the hotter it gets, which is exactly backwards and flatters the result.
• Constant current. As resistance rises, voltage rises, and power rises. The thermal load runs away from the setpoint, and the test is harsher than the specification in an uncontrolled way.
• Constant power. The supply trades voltage against current to hold the product constant. The thermal load stays at the rated wattage, which is the quantity the cooling module’s specification actually describes, so the element’s temperature coefficient stops mattering.

This is why an AI server cooling bench runs the source in constant power (CP) mode rather than in the C.V or C.C priority modes used for most supply-voltage work. Both the PHU Series and the PSU Series run CP mode for thermal loading.

The air-cooled bench

The documented configuration uses a PHU 1500-30 delivering 1 kW in constant-power mode into a CPU simulator. The simulator sits under the cooling module being evaluated, and the assembly goes into a wind tunnel where airflow is controlled, so that cooling efficiency, temperature distribution and overall thermal performance are measured against a known thermal input and a known airflow.

Two variables have to be pinned for the result to mean anything, and the wind tunnel exists to pin the second one:

• Thermal input, held by the supply in constant-power mode.
• Airflow, held by the tunnel, because a heat sink’s performance is a function of the air moving through it and an open bench does not control that.

Take the wattage from the device you are actually cooling. The 1 kW figure is the one in the source material, not a universal AI server figure, and it is worth noting that it sits well above the TDP of any shipping server CPU, which currently tops out around 500 W. A 1 kW thermal load is an accelerator, a module, or a multi-device envelope.

The liquid-cooled bench

Liquid cooling changes the hardware around the load, not the principle. The thermal simulator mounts to the cold plate instead of sitting under a heat sink, and the measurement moves to coolant inlet and outlet temperatures, flow rate and pressure drop alongside the simulator’s own temperature.

The configuration for this case is a PSU Series PSU 300-6 in constant-power mode delivering 240 V at 6 A, which is 1440 W into the thermal load, with two or three units paralleled where the cooling system’s rated load exceeds what a single unit delivers. That parallel step matters because a PSU Series unit is rated at 1200 W to 1560 W, so a loop designed to remove several kilowatts needs it from the start. The series permits up to four units of the same model in parallel.

Size the supply from the loop’s rated thermal load and the simulator’s resistance rather than from the example.

What to log, and for how long

Thermal results drift. A cooling module that looks fine in the first ten minutes can fail at the second hour as the thermal interface material settles, the coolant warms, or a fan curve shifts. The measurement is therefore a long-duration one, which is what the DAQ-9600 is on the bench for: multi-channel thermocouple logging across the simulator, the module, the inlet and outlet, and ambient, for the full test duration rather than a spot reading.

Log at minimum:

• Simulator power, as delivered, not as set. Confirm the supply held the setpoint.
• Simulator case temperature, and the module’s own temperature at the points the specification names.
• Ambient and, for liquid, coolant inlet and outlet temperatures and flow rate.
• Airflow for an air-cooled test, including whether the tunnel was holding it steady.

The pass criterion is a temperature rise at a stated power, so a result missing either half of that pair is not reportable.

Where this sits in an Indian programme

Thermal module work in India sits mostly with heat sink, cold plate and thermal solution suppliers building for server and AI hardware programmes, and with the integrators qualifying what those suppliers deliver. It is design-verification work that turns into sampled production verification, and the bench is shared: the same programmable source and DAQ that run a thermal qualification also run power-supply and board-level work the rest of the week.

For the wider AI server power picture, see AI data center testing: power shelves, busbars, burn-in and optics.

Instruments and support from GSAS

GSAS Micro Systems is an authorized GW Instek engineering partner for Karnataka, Andhra Pradesh and Telangana, supporting customers across India. We supply the PHU Series, PSU Series and DAQ-9600 with application engineering on thermal-bench architecture and instrument sizing, and calibration support, from Bengaluru, Chennai, Pune, Hyderabad, Mumbai and Delhi NCR.

To scope a thermal bench against your own cooling module and its rated load, talk to us.

Source: GW Instek application material for the PHU Series and PSU Series. The 1 kW air-cooled figure is the one in that material; thermal input for your own test comes from the device being cooled.