PCB Crosstalk: NEXT, FEXT, and Mutual Coupling

Why a neighbouring trace picks anything up at all

Crosstalk is your neighbour's signal leaking onto yours. You route two traces side by side — a fast data line and a quiet analog reference, say — and when the data line switches, the analog line twitches. No physical connection exists between them, yet energy has crossed over. On a tight-pitch bus, every line is both aggressor and victim at the same time, and the coupled noise can close an eye as effectively as loss or reflections.

The mechanism is not mysterious once you see it: any two conductors near each other share electric and magnetic fields. A changing voltage on one pushes current through the shared capacitance. A changing current on one induces a voltage through the shared inductance. These are the same i = C·dV/dt and v = L·di/dt from a first circuits course, just happening between traces instead of inside a component.

Mutual capacitance. Two conductors near each other form a capacitor, even when neither is supposed to be one. A changing voltage on the aggressor pushes current through that capacitance and into the victim — the familiar i = C·dV/dt. Note what that says: the coupling depends on how fast the voltage is changing, not on how often.

Mutual inductance. The aggressor's current creates a magnetic field, and part of that field passes through the loop formed by the victim and its return path. A changing field through a loop induces a voltage — v = L·di/dt. Again it is the rate of change that matters.

Both coupling mechanisms are present, but their relative strengths depend on geometry and dielectric distribution. The terms do not point the same way, which produces different near-end and far-end behavior.

Two couplings, two directions

Here is the asymmetry that explains everything else on this page. The capacitively injected current has no preferred direction — it splits evenly and travels both ways along the victim. The inductively injected voltage does have a direction, set by the sign of the changing field: it drives current backwards, towards the near end.

So at the near end the two contributions add, and at the far end they subtract.

K_b = ¼ ( L_m/L₀ + C_m/C₀ ) K_f ∝ ( L_m/L₀ − C_m/C₀ ) K_b is the backward (near-end) coefficient, K_f the forward (far-end) one. The sign between the two terms is the whole story.

Set the two ratios equal on the panel above and watch the far-end pulse vanish while the near-end plateau stays put. That is not a numerical coincidence — it is the subtraction going to zero, and it is the reason a buried stripline behaves so differently from a surface microstrip.

NEXT saturates. FEXT accumulates.

Near-end crosstalk is a low, wide pulse. Each slice of coupled length contributes noise that arrives back at the near end at a slightly different time, so the contributions smear into a plateau lasting twice the coupled delay. Crucially, its amplitude stops growing once the coupled length exceeds t_r / 2t_pd — past that, extra length makes the pulse longer, not taller. For a 100 ps edge on stripline, that is about 0.3 inch. Runs longer than this example's saturation length mainly extend the NEXT pulse duration rather than its height, which is why NEXT is often quoted as a percentage of aggressor swing.

Far-end crosstalk behaves the opposite way. It arrives as a single narrow pulse about one rise time wide, travelling alongside the aggressor edge and gathering contributions the whole way — so its amplitude grows linearly with coupled length and gets worse as edges get faster. It lands at the receiver at exactly the moment the data edge does, which is what makes it a timing problem rather than just a noise problem.

Stripline versus microstrip In a homogeneous dielectric — a buried stripline — the inductive and capacitive coupling ratios are equal, so the far-end terms cancel and FEXT is ideally zero. A microstrip has air above and laminate below, the ratios no longer match, and FEXT appears. This is the strongest SI argument for burying a congested bus: it can greatly reduce this FEXT mechanism. In practice a real stripline still shows some FEXT, mostly from glass-weave inhomogeneity and asymmetric builds — but far less than the equivalent microstrip.

The levers, in the order they are worth pulling: increase spacing (coupling falls roughly with the square of separation, so 3× trace width buys most of what is available), reduce the coupled length, move the pair closer to its reference plane — halving the dielectric height roughly quarters the coupling — or bury the run. Guard traces are a distant last resort and are actively harmful unless stitched to ground densely along their whole length.

Why this decides how your board behaves

The two kinds of crosstalk fail in different ways, and they want different fixes.

NEXT is a noise problem at the driver's end. It is a long, low plateau, so it looks like a shifted reference rather than a glitch. On a bidirectional bus it lands on the receiver that is currently listening while its neighbour drives — which is why NEXT matters most on wide parallel buses that turn around.
FEXT is a timing problem. It arrives as a narrow pulse at exactly the moment the victim's own edge does, because both travelled the same distance at the same speed. It moves the crossing, so it shows up as jitter rather than as noise, and it is worse on longer runs and faster edges.
It is not one aggressor. In a bus, a victim has neighbours on both sides and further out. Contributions add, and the worst case is all aggressors switching the same way at once — which is exactly what a wide data bus does.
An equaliser makes it worse, not better. Crosstalk lives at high frequency, and a CTLE boosts high frequency. It lifts the crosstalk along with the signal. See what closes the eye, where you can turn aggressors up with the equaliser on and watch it happen.

What to do about it

The levers, in the order they are worth pulling.

First, spacing. Coupling falls roughly with the square of separation, so the first doubling buys most of what is available. Three times the trace width between centres is the usual starting rule, and going beyond about five buys very little. This is the cheapest effective when the required spacing fits the routing constraints.

Second, get closer to the reference plane. Halving the dielectric height roughly quarters the coupling, because the field that would have reached the neighbour terminates on the plane instead. It also halves the trace width needed for the same impedance, which gives back some of the space spacing cost you. This is a stackup decision and often a better one than a routing decision.

Third, shorten the coupled length — while noting which component dominates. Past saturation it may not reduce the ideal NEXT plateau height, although it shortens the pulse; in the weak-coupling model it reduces FEXT approximately in proportion to length.

Fourth, consider stripline for a congested run. A homogeneous stripline can cancel the ideal far-end term, while manufacturing asymmetry and transitions still contribute residual coupling. Evaluate the complete routed structure before assigning the benefit.

Termination and timing complete the result Coupling coefficients describe the interconnect, while the voltage observed at the victim also depends on source and load impedances, transition direction, relative timing, and the states of other aggressors. Preserve those conditions when comparing a field-solver result, a transient simulation, and a bench waveform.

Guard traces are a distant last resort. An unstitched guard trace is a floating conductor that couples to both neighbours and can make things worse. If you use one, it must be stitched to ground densely along its whole length — vias every fraction of a wavelength at the knee frequency — and at that point you have usually spent more space than simply separating the two signals would have cost.

Using this to find a fault

Noise on a quiet net, lasting much longer than an edge. A wide, flat plateau is NEXT. Measure its duration: it should be twice the coupled delay, which tells you how long the shared run is and therefore where it is.
Jitter that appears only when neighbours are active. FEXT is one candidate. It should correlate with the aggressor's data pattern, not with anything on the victim, which is the quickest way to tell it apart from ISI.
Crosstalk between traces that are not adjacent. Look at the return paths, not the traces. Two signals whose returns are forced through the same plane opening are coupled regardless of how far apart the traces run. See return current paths.
A bus where only some bits fail. Map the failures against physical position. Edge bits have one neighbour, middle bits have two — if the middle of the bus fails first, crosstalk is the likely cause; if the edges fail first, look at the return path at the edge of the field instead.

Go deeper — where the coefficients come from, and why saturation happens

Take a short slice of coupled line, length δ. The aggressor edge passing over it injects a small amount of noise into the victim: a capacitive component that splits both ways, and an inductive component that goes backwards only. Integrate those contributions over the whole coupled length and the two behaviours fall out of the geometry of when each contribution arrives.

Backward. Noise generated at distance x along the coupled section travels back to the near end, arriving at time 2x/v after the edge set off. Different slices therefore arrive at different times, spreading the contributions into a plateau lasting 2T_d. Because they are spread in time rather than stacked, the height of the plateau is set by the coupling per unit length and the rise time — not by the total length. Once the coupled section is longer than the edge occupies physically, adding more only makes the plateau wider:

l_sat = t_r / ( 2 · t_pd ) for a 100 ps edge on stripline, about 0.3 inch; longer runs are in the saturated-height region of this ideal NEXT model

Forward. Noise generated at x travels forward at the same speed as the aggressor edge that made it. It never falls behind, so every slice's contribution arrives at the far end at the same instant and they stack directly on top of each other. That is why FEXT grows linearly with length and NEXT does not — and why FEXT is a narrow pulse about one rise time wide rather than a plateau.

The cancellation is then easy to state. The forward term is proportional to (L_m/L₀ − C_m/C₀). In a homogeneous dielectric — everything buried in the same laminate — the inductive and capacitive coupling ratios are forced to be equal, because both are determined by the same field geometry and the same material everywhere. The difference is zero and FEXT vanishes. Put air above the trace and the electric field, which is happy in air, behaves differently from the magnetic field, which barely notices the material. The ratios separate and FEXT appears.

Two honest caveats on the model in the panel. It is a weak-coupling superposition, valid while the ratios are below roughly 0.3 — beyond that the aggressor's own impedance is changed by the presence of the victim and the linear treatment breaks down. And it is a single pair; a real bus needs the contributions from several aggressors summed, with the worst case being all of them switching together in the same direction.

In the real world

The practical shape of crosstalk work is a budget rather than a universal spacing rule. Routing space is limited, so the job is deciding where the space goes: clocks and strobes get isolation, address and control get whatever is left, and the decision is made by what a coupled pulse would cost at each receiver rather than by a uniform spacing constraint.

The failure worth anticipating is the one that only appears in the assembled system. Crosstalk inside a connector, between pins of a package, or across a cable bundle is often larger than anything on the board. These structures deserve explicit review even when they arrive inside a model supplied by another team or vendor.

Return-Current Paths and Reference Continuity
Differential Signaling and Mode Conversion
PCB Stackup Design and Dielectric Selection
Simultaneous Switching Noise and Ground Bounce
Lab D: ADC Interference, Aliasing and Reference Noise — the same dV/dt coupling, arriving at a converter’s input or its reference