How a common-mode choke works, and where it goes on USB

The decision rule first: a common-mode choke is the right part when the noise is common mode — the same voltage, in phase, on both wires of a pair with respect to ground — and the signal is differential. Its coupled windings present a high impedance to the first and, ideally, none to the second, which is something no combination of ordinary inductors and capacitors can do. If the noise is differential, a choke is the wrong part; if the signal is single-ended, it is the wrong part twice.

And a choke is selected from two curves, not one number. The common-mode impedance curve must be high across the band where the noise actually is, and the differential-mode (leakage) curve must be low across the band where the signal is. The 100 MHz figure printed on the datasheet is one point on the first curve and says nothing about the second — which is the curve that decides whether a 480 Mbit/s eye survives the component.

What a common-mode choke is

A common-mode choke is two windings on one magnetic core, one in each wire of a pair, wound so that a current flowing the same way in both wires magnetises the core and a current flowing opposite ways in the two wires does not. The signal on a pair is the second kind: it goes out on one wire and comes back on the other, its two fields cancel in the core, and the choke is close to invisible to it. Noise picked up from the environment, or pushed onto both wires by a switching supply, is the first kind: it flows the same way in both, its fields add, and the choke presents it with the full inductance of the winding. That is the entire mechanism, and everything below is about how well a real part does it and where it stops.

Two currents on one pair of wires

Every two-wire interface carries two superimposed circuits. The differential current — the signal — flows out on one wire and back on the other; its loop is the pair itself, and a receiver that measures the voltage between the wires sees all of it. The common-mode current flows the same direction on both wires at once, as if the pair were a single fat conductor, and has to close its loop somewhere else: through the ground system, the chassis, or parasitic capacitance to earth.

That second loop is the problem. The receiver rejects the common-mode voltage by construction — the same property that lets an RS-485 bus run through an electrically hostile building — so common-mode noise rarely corrupts the data directly. What it does instead is radiate, because its return loop is large, ill-defined, and usually includes the cable.

Why an ordinary filter cannot remove it

Würth ANP146 opens with an LTspice experiment that makes the case better than any argument: a 1 V differential signal at 500 kHz and a 1 V common-mode interferer at 5 MHz on the same pair. A single series inductor with a capacitor — the ordinary LC low-pass — in one leg attenuates only the differential signal and leaves the common-mode noise untouched, because the noise is on both legs with respect to ground and an unbalanced filter only touches one. Putting an independent choke in each leg does attenuate the noise, by about 30 dB in the simulation — but it strongly attenuates the wanted signal too, since each inductor is in series with it.

The way out is what Würth call current compensation: wind both inductors on the same core, with the right sense.

Coupled windings sort the currents by mode

In a common-mode choke, the windings act in such a way that the magnetic fields of the currents in differential mode cancel each other out, while they add up in common mode. The current in differential mode flows through the windings in opposite directions, whereby the magnetic fields generated ideally compensate each other completely. As a result, the inductance for differential currents approaches zero, which does not attenuate the normal signal flow. With common mode currents, the currents flow in the same direction through both windings, whereby the magnetic fields are added together, and a high inductance is created.

The component is a transformer used sideways. How well the trick works is set by the mutual inductance M between the windings, expressed as the coupling coefficient:

$K = \frac{M}{\sqrt{L_1 \cdot L_2}}$

K runs from 0 to 1, and ANP146 gives 0.80 to 0.95 as the typical range for common-mode chokes. In the same LTspice setup, replacing the two independent inductors with a coupled pair leaves the 500 kHz signal unattenuated while the 5 MHz common-mode interferer drops by roughly 40 dB. Same two inductors, same core material — the winding sense is the entire difference.

The noise is trying to leave on the cable

Where does the common-mode current come from in the first place? On a USB port, mostly from inside the box. Würth ANP024 traces it to parasitic coupling in the circuit environment of the controller — usually capacitive, in its words, and growing with frequency; a switching converter or a clock next to the data lines is the typical source of that coupling. It lands on D+ and D− together, in phase and at the same amplitude, so the differential receiver never notices. The cable does: driven in phase against earth, it is no longer a transmission line but an antenna, and ANP146 describes a peripheral cable in exactly those terms — a monopole antenna with high impedance when mismatched.

The same mechanism runs in reverse, and worse:

Asymmetries in the cable or at the receiver often convert the originally common mode into a differential mode interference signal, however, which can then contribute to signal impairment.

Mode conversion is why “the receiver rejects common mode” is not the end of the story. Any imbalance — an unmatched pair, one leg loaded more than the other, a cheap cable — converts some of the harmless common-mode noise into differential noise sitting directly on the signal. The choke’s job is to stop the common-mode current before anything gets the chance.

The impedance curve is the component

ANP146 works through a real part, the WE-CMDC 744238132, and the exercise is worth following because it shows how little the headline number describes. The datasheet gives a scalar impedance of 910–1300 Ω at 100 MHz and a copper resistance of at most 25 mΩ per winding. Measuring the equivalent circuit gives the shape behind that point: a total inductance L_0 of 18.91 µH (12–13 Ω of impedance at 100 kHz, phase 82°), a common-mode curve with a flat, lossy maximum around 100 MHz — the ferrite is dissipative across the whole band, so the resonance is barely pronounced — and a differential-mode curve with a sharp resonance, measured at 241.77 MHz, caused by the parasitic capacitance of the windings. ANP146 puts that capacitance at 1.26 pF “including all measurement stray effects as well” — a figure that does not square with its own resonance, since 2.01 µH and 1.26 pF would resonate near 100 MHz. The curve below uses the value that follows from the two measured numbers, about 0.22 pF; the shape is the same, and the resonance sits where the network analyzer put it.

Read as a pair, the curves say what the part does: an order of magnitude more impedance for the common mode than for the differential mode below a few megahertz, where both windings are still simply inductive — about 9:1 at 1 MHz in the model, narrowing to 7:1 at 10 MHz and 3:1 at 30 MHz as the differential curve climbs toward its resonance. Read alone, “1 kΩ at 100 MHz” says almost nothing — the same lesson as a ferrite bead’s rated impedance, and for the same reason: one point on a frequency-dependent, lossy curve is not a specification.

Leakage inductance: the differential part you get anyway

K never reaches 1, so some of each winding’s flux fails to link the other winding. That unlinked remainder is the leakage (stray) inductance L_S, and it is in series with the differential path — the signal path. ANP146 measures it by shorting the secondary, which cancels the coupled flux and leaves only the stray part:

$L_0 = L_S + L_0'$

For the 744238132 the shorted measurement gives L_S = 2.01 µH, and from it the coupling:

$$K = \frac{L_0 - L_S}{L_0} = \frac{18.91 - 2.01}{18.91} = 0.89 \qquad M = K \cdot L_0 = 16.9\ \mu\text{H}$$

Every common-mode choke has this second, unavoidable component. The only question is whether 2 µH in series with your signal is a gift or a defect, and the answer depends entirely on what the line carries.

Feature or bug: two winding styles, two answers

ANP146 is explicit that the trade is a design parameter, not an accident:

A high K means strong coupling and minimal stray inductance, which reduces the choke’s impedance to DM currents, lowering its DM attenuation. A lower K corresponds to higher stray inductance, increasing DM impedance and improving DM noise suppression. However, a high K is critical for maximizing CM attenuation, as it increases the choke’s impedance to CM currents.

On a supply line, leakage is a feature. There is no fast signal to protect, and 2 µH of stray inductance plus an X-capacitor is a free differential-mode filter riding on the same component. ANP146 ties that trade to the winding style: with sectional winding, where the two windings occupy separate sections of the core, K can be adjusted to balance the two attenuations — the measured WE-CMDC is sectional, and the note later uses it in a 24 V DC input filter. With bifilar winding K is inherently high, so the trade-off does not apply and differential-mode noise needs an additional measure — a lossy SMD ferrite bead, a component with selection rules of its own. On a 480 Mbit/s data line the same microhenries are a bug: they sit in the signal path as a plain series inductor, and the leakage budget is nanohenries. That is bifilar territory — the WE-CNSW that ANP002 and ANP024 put on USB is the bifilar example in ANP146’s own Figure 1.

Y-capacitors turn the choke into a divider

A series impedance only attenuates relative to what follows it, and what follows a common-mode choke is often the worst case: a mismatched cable, high impedance, effectively unloaded. ANP146 states the fix as a requirement:

To achieve sufficient attenuation even for high-impedance CM loads, such as peripheral cables (→ monopole antenna with high impedance in the event of mismatch), the output of the filter must be low impedance in terms of RF technology.

Capacitors from each line to ground — Y-capacitors — on the output side of the choke give the divider a low-impedance bottom leg, exactly like the shunt leg of a first-order RC filter. In the ANP146 simulation this takes the same choke from roughly 40 dB of common-mode attenuation to over 70 dB, and makes the attenuation independent of the load impedance, because the division now happens between the choke and the capacitors rather than between the choke and whatever the cable feels like being today.

What each mode pays through the filter

ANP024 puts numbers on the asymmetry for the USB data-line choke it designs around (the 744232090): common-mode impedance about 90 Ω at 100 MHz, differential-mode impedance 6 Ω at the same frequency for both windings together — fifteen to one in favour of the signal.

It is its common mode or asymmetrical suppression that takes effect if the same interference components are on D+ and D- with respect to ground. This is always the case for capacitive or inductive coupling on the circuit or its conductor tracks. So this impedance component must be as high as possible.

With the TVS array’s 6.6 pF doing duty as the shunt capacitance, ANP024’s divider arithmetic gives about 10.6 dB of common-mode attenuation at 100 MHz for the complete data-line filter, and a matched 50 Ω measurement — the floor, computed in the figure above — shows only a few decibels there. Yet the measured interference spectrum on D+ of a real USB hard drive drops by up to 35 dB behind the same filter: ANP024 notes the suppression depends strongly on the impedance of the interference source, and the parasitic coupling that puts noise on a data line is high-impedance. The differential signal, meanwhile, pays a fraction of a decibel everywhere. That asymmetry — not any absolute number — is what the component is for.

USB 2.0: the order of parts

Würth ANP002 gives the arrangement, and notes that Intel’s High Speed USB Platform Design Guidelines already recommend both parts: a common-mode choke for EMI and a separate component for ESD. The recommended routing runs the two data lines from the connector to the TVS diode array first, then through the common-mode choke to the controller — the strike is clamped at the board edge, before the choke, and the TVS array’s sub-2 pF line capacitance is small enough to be nearly invisible to the signal.

VBUS is routed the same way but gets a chip ferrite bead rather than a choke — a single supply line has no differential pair to preserve, so a lossy series impedance is the right filter — with the option of a capacitor and a second bead for more suppression. And ANP002 flags the omission that actually sinks boards:

But for outstanding EMI behaviour it is necessary to protect the power supply (VBUS) as well. A lot of design engineers forget this important point and wonder themselves when their product doesn’t pass all the tests in EMI lab.

A perfectly filtered data pair next to an unfiltered VBUS wire in the same cable is not a filtered port; the noise simply takes the wire without a choke on it.

Shield ground is not signal ground

The filter’s shunt elements — Y-capacitors, TVS diodes — only work if their ground is quiet at RF, and the cable shield only works if it is bonded to the same potential. ANP024’s layout rules are blunt: a capacitor’s ground connection through a centimetre of track is 6–10 nH in series with it, and

The ground connection should pass directly to the housing, as the ground reference of the cable shielding and the ground reference of the filter must lie on the same HF potential.

Bond the connector shell to chassis at the connector, and give the filter its ground reference at the same place. The cable itself is part of this chain: ANP024 specifies a braid of at least 65 % coverage plus an aluminium foil, worth better than 50 dB of shield attenuation — and notes that cheap cables with a one-sided metallised foil manage barely 15 dB. A choke cannot fix a cable whose shield is decorative.

What excess leakage does to the eye

USB 2.0 high speed runs at 480 Mbit/s with a 400 mV signal into 45 Ω terminations on each line — 90 Ω differential — so a bit lasts about 2.08 ns. The choke’s leakage inductance is in series with that signal, between the 90 Ω source termination and the 90 Ω load termination, so the L/R time constant that slows the edge is L_S/180 Ω.

The 6 Ω differential impedance of the USB-rated choke corresponds to under 10 nH of leakage: a time constant of ~0.05 ns, invisible inside a 2 ns bit. The 2.01 µH leakage of the sectional-wound WE-CMDC part measured in ANP146 gives 11 ns — more than five bit times — and the edge simply never arrives. This is the whole reason data-line chokes are specified per interface and per data rate: the common-mode side scales freely, but the leakage and parasitic capacitance budgets are set by the eye mask. Capacitance obeys the same budget — ANP024 tolerates about 3 pF from each data line to ground before USB 2.0 distortion sets in, and ANP002 puts the practical limit for passing the eye-pattern test near 5 pF. A choke that closes the eye has not filtered the interface; it has become it.

Choosing the part, in order

1. Establish that the noise is common mode, and find its band. A choke does nothing for differential noise, and a big inductance does nothing for noise two octaves above the impedance peak.
2. Data line or supply line? A fast data line needs high K and nanohenries of leakage — a bifilar part rated for the interface and its bit rate; a supply line or a slow link can use a sectional part, whose adjustable K leaves leakage to double as the differential-mode filter.
3. Read the common-mode curve across the noise band — the 100 MHz number is one point, and the curve around it is lossy, flat, and finite.
4. Read the differential-mode curve across the signal band, because the signal pays that impedance in edge rate.
5. If the load impedance is high or unknown, add Y-capacitors and let the divider, not the load, set the attenuation — with their grounds and the shield on one HF potential at the connector.

Sources

• Würth ANP146 — Theoretical Insights and Practical Applications of the WE-CMDC Series Common Mode Chokes — the current-compensation mechanism, the coupling coefficient, the measured equivalent circuit of the 744238132, the Y-capacitor divider, and the leakage-as-DM-filter trade.
• Würth ANP002 — The Protection of USB 2.0 Applications — the connector → TVS → choke → controller order, TVS capacitance limits, the eye-pattern test, and the VBUS warning.
• Würth ANP024 — The USB Interface from EMC Point of View — mode conversion, the data-line filter design and its measured 35 dB, the 90 Ω/6 Ω choke, layout inductance, and shield grounding.

Updates

• 2026-09-21 — Added a plain statement of what the part is, ahead of the decision rule, for readers arriving with that question rather than a USB layout.
• 2026-09-13 — Fig 9 and the eye section: the leakage time constant in a doubly terminated USB pair is L/180 Ω, not L/90 Ω, since the series inductance sits between the 90 Ω source and 90 Ω load terminations. 2.01 µH gives 11 ns (about five bit times), not 22 ns; the 100 nH curve redrawn accordingly.
• 2026-09-13 — The “ten to one” between the common-mode and differential-mode curves holds below a few MHz; the ratio narrows to about 3:1 by 30 MHz as the differential curve climbs toward its resonance. Text and Fig 4 now say so.
• 2026-09-13 — Fig 4 states the parasitic capacitance it uses (≈ 0.22 pF, which puts 2.01 µH at the measured 241.77 MHz); ANP146’s 1.26 pF, which it qualifies as including measurement stray effects, would resonate near 100 MHz.
• 2026-09-13 — The 744238132 is no longer called a “power-line choke”: ANP146 files the sectional WE-CMDC series under data-line filters and uses it in a DC input filter. The winding-style section now states what the note says — sectional winding lets K be adjusted, bifilar winding has inherently high K and needs a ferrite bead for differential-mode noise — rather than a sectional-is-power, bifilar-is-data rule.
• 2026-09-13 — Fig 7 redrawn so the D+ stub to the TVS array hops over D− instead of appearing to short to it; attribution of the coupling mechanism and the “monopole” description tightened to the sources’ wording.