GPU Power Board Testing: DC Input Verification Before Integration

The economics of this test are simple. A GPU power delivery board that fails on a bench costs a board. The same board failing inside an assembled AI server costs the teardown, the retest, the schedule, and the chance that the fault is blamed on something else entirely before anyone looks at the power stage.

So the board gets a DC input test before it goes anywhere, and the same discipline gets applied to whole systems on arrival.

What the test establishes

A programmable DC source supplies the board’s input while the test system monitors input voltage and current, power consumption, output behaviour and protection response, at operating points covering start-up, normal operation and higher-power conditions.

The value is in the plural. A single operating point tells you the board works at one condition. A programmed set of them tells you how it behaves across its range, and failures cluster at the edges:

• Start-up. Inrush into the board’s input capacitance, the sequencing of its output rails, and whether the controller comes up cleanly from a cold start.
• Normal operation. Conversion behaviour and current draw at the condition the board will spend its life in.
• Higher-power conditions. Where thermal and control-loop margin runs out, and where a marginal component separates from a good one.
• Protection. Whether the board’s own protection trips where it should, and recovers how it should.

Repeatability is the whole point

A one-off measurement on a board tells you about that board. A production screen has to tell you about the difference between boards, and that only works if everything except the board is held still.

That is the argument for a programmable supply over a bench supply with knobs. The sequence of operating points, the dwell at each, the ramp between them and the protection limits are all set in the instrument and reapplied identically to every unit. When a board reads high on input current, you know it is the board.

The PSU Series suits this role because of what it is rather than anything exotic: 1200 W to 1560 W per unit in a 1U chassis across fifteen voltage ratings from 6 V to 600 V, with C.V and C.C priority, adjustable voltage and current rise and fall times, and RS-232, RS-485, USB and LAN for automation. The adjustable rise time matters more than it looks, because a board’s start-up behaviour depends on how fast its input comes up, and a test that does not control that is not controlling its own stimulus.

The same test, one level up: incoming system inspection

GW Instek’s application material applies the identical approach to whole AI server systems on arrival, before they enter integration, burn-in or production test. The source supplies the specified input condition, and the test system monitors input current, power consumption, start-up behaviour and basic system operation, with power-on, standby and operating conditions repeated consistently for every unit.

The rationale is the same as the board case, scaled: finding an abnormal current draw or a start-up failure at goods-inwards is cheap, and finding it after the unit has been integrated is not. It is screening, not characterisation, and it should be fast enough that it does not become the bottleneck it was meant to prevent.

Setting the limits

None of the numbers in this test come from the instrument. The input voltage, the operating points, the expected current at each, and the pass band around it all come from the DUT specification, and GW Instek’s own material says so explicitly.

Two practical notes on getting those limits right:

Set the pass band from measured spread, not from the nominal. Run a sample of known-good boards first and set limits around what they actually do. A band derived from the design nominal will either pass everything or fail good units, usually both at different operating points.

Record the condition, not just the verdict. A board that passed is less useful in a field-failure investigation than a board that passed with its input current logged at each point. The logging is nearly free once the sequence is automated.

Where this sits in an Indian programme

This is contract-manufacturing and systems-integration work, and it is growing in India with server and AI hardware assembly. It is normally the first electrical test in the flow, which gives it an outsized effect on everything downstream: a screen that is too loose pushes faults into integration, and a screen that is too tight stalls the line.

The bench is shared with neighbouring work. Board-level verification and incoming inspection run on the same supply, the thermal loading described in emulating processor power for thermal testing is the adjacent job on the same bench, and the wider rack and power-shelf picture is in 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 PSU Series, PEL Series electronic loads and the DAQ-9600 with application engineering on test sequencing and limit setting, and calibration support, from Bengaluru, Chennai, Pune, Hyderabad, Mumbai and Delhi NCR.

To scope a board or incoming-inspection bench against your own specification, talk to us.

Source: GW Instek application material for the PSU Series, which states that exact voltage, current, power and acceptance limits are defined from the customer’s DUT specification.