Testing an EMI filter feedthrough capacitor comes down to two levels: measure the part’s insertion loss on a 50 Ω bench setup, then prove the whole PSU passes conducted emissions on the LISN. If you only check capacitance with a meter, you know the part is alive—not that it filters. This article walks through both levels for AI server power supplies, plus the four failure patterns we see most often when a filter underperforms on the bench.

Two Levels of Verification: Component and System

A feedthrough capacitor passes qualification in two places, and they answer different questions. The component-level test answers “did the part come out of production as designed?” The system-level test answers “does the finished power supply meet CISPR 32 Class B at the rack?” At Xuansn, both gates run on every production batch before parts ship.

• Component level — insertion loss per CISPR 17, plus capacitance, dissipation factor, insulation resistance and hi-pot. This is the manufacturer’s production screen and your incoming inspection.
• System level — conducted emissions from 150 kHz to 30 MHz measured through a LISN, per CISPR 32. This is what your compliance lab actually signs off.

Confusing the two is the most common workflow mistake we hear from engineering teams. A part can show 40 dB of insertion loss on the spec sheet and a supply can still fail conducted emissions—because layout, source impedance and cable radiation dominate at those frequencies, not the filter alone. Treat the two tests as sequential gates, not substitutes.

Measuring Insertion Loss per CISPR 17

CISPR 17 is the standard most manufacturers reference for filter insertion loss. The classic method uses a calibrated signal source and receiver in a 50 Ω system, measuring the ratio of received level without the filter to received level with the filter in circuit. Because real power lines are not 50 Ω, the standard also defines asymmetric test fixture impedances (0.1 Ω/100 Ω combinations) to approximate actual line conditions—so ask your supplier which mode the datasheet numbers use.

On your own bench, a practical routine looks like this:

• Calibrate the analyzer path first (through calibration), so what you record is the filter, not cable loss.
• Mount the part the way it will be used: bulkhead or panel-mounted against a ground plane, not floating on leads in the air. A feedthrough part measured with long leads loses its coaxial geometry—and most of its high-frequency performance with it.
• Read insertion loss across the band you actually need to suppress: the converter switching frequency and its harmonics, typically from a few hundred kilohertz up to 30 MHz and beyond.
• Repeat at high temperature and after humidity or load cycling if the program requires it—insertion loss drifts when the dielectric or termination degrades.

For AI PSU work, we suggest reading the curve at the switching frequency ± harmonics rather than trusting a single “minimum insertion loss” number at an unspecified frequency.

Whole-PSU Verification Against CISPR 32

Component data gets you shortlisted; the compliance lab decides. For conducted emissions, the finished PSU connects to a line impedance stabilization network, and a receiver scans 150 kHz–30 MHz against CISPR 32 limits—Class A for equipment in industrial environments, Class B for residential. AI servers deployed in enterprise data centers are usually held to Class A, but check the program’s contractual requirement first; some operators impose Class B quietly.

What the feedthrough contributes here is measurable: compare the emissions scan with and without the filter stage populated. The delta at the switching-frequency peak is the filter’s real-world contribution, and it is the number to negotiate over when a customer pushes back on cost.

When the Numbers Disappoint: Four Field Fixes

The curve looks great in the datasheet band but collapses above 10 MHz

Almost always a mounting problem. Lead inductance in series with the feedthrough destroys its coaxial bypass path. Fix: mount directly against the chassis ground plane, thread-tightened or soldered per the manufacturer’s torque/solder profile, with the input and output sides physically separated.

Conducted emissions fail at the switching frequency fundamental

The filter capacitance is too small for the source impedance the converter presents, or the resonance between the filter and the input impedance of the downstream converter sits right at the fundamental. Fix: revisit the capacitance value against the actual measured noise spectrum (our parameter design article covers the LC math), or add a small series inductor to shift the corner.

Passes CISPR 32 on a clean bench, fails in the rack

Rack conditions add cable coupling, adjacent-slot sources and a shared ground that a single-PSU bench never shows. Fix: reserve margin of at least 6 dB below the limit line on the bench, and verify inside a powered rack or with a chassis-grounded harness before committing the design.

Insertion loss drifts after thermal cycling

Termination degradation or moisture ingress. Fix: require hi-pot and IR re-test after cycling from your supplier, and specify the endurance grades you need rather than assuming all parts of the same capacitance behave the same. For AI PSUs running hot at the rear panel, this is the failure mode that shows up six months after shipment.

Acceptance Criteria Worth Writing Into Your Spec

When AI server programs qualify a feedthrough filter stage, a short written acceptance table prevents most of the disputes we see between OEM and supplier:

Numbers in the “typical acceptance” column are engineering starting points, not universal limits—align them with your program’s safety file and the supplier’s qualification data.

Common Questions About Testing Feedthrough Capacitors

Can I test insertion loss with a VNA instead of CISPR 17 fixtures?

Yes—a two-port VNA with S21 through-measurement gives the same 50 Ω insertion-loss information and is what most labs actually run day to day. Just calibrate out the fixture, keep the part mounted against a ground plane, and note that VNA results below roughly 100 kHz get harder to trust as the fixture dominates.

What margin should I design in before the compliance lab visit?

Six dB below the CISPR 32 limit at every peak is a workable rule of thumb for AI PSUs. Labs, cables and racks all add a few dB of variance, and re-testing after a layout spin is far more expensive than the margin was.

Does the supplier’s datasheet insertion loss match what I will measure?

Only if your mounting matches theirs. Datasheet curves assume the fixture geometry in the test report—panel-mounted on a defined ground plane. Bulkhead installations with paint, anodizing or gaskets between part flange and chassis will measure worse, so budget for it in the margin.

This is the third part of our AI power EMI series. Start with why AI server PSUs need EMI filtering and how to select an EMI filter feedthrough capacitor, or read the feedthrough capacitor fundamentals guide for the underlying physics. For datasheet-level detail, see our manufacturing process article.