Crosstalk: Mutual Capacitance, Mutual Inductance, and Coupling Length

Two traces running side by side share the space between them. A field launched on one induces a current in the other, and the induced current appears at the victim’s receiver as noise. The effect is unavoidable in any parallel routing, so the design question is not how to eliminate coupling but how to keep the coupled amplitude below the receiver’s noise budget — and how to make sure the coupled noise arrives at a time when the receiver is not sampling.

Two mechanisms, two polarities

Coupling between parallel traces has a capacitive and an inductive component, and they do not add with the same sign at both ends.

The capacitive part is straightforward: the aggressor’s voltage drives current through the mutual capacitance between the traces into the victim. The injected current splits at the victim’s own impedance and produces a voltage. This current is proportional to $dV/dt$ and to the mutual capacitance, which grows as the traces get closer and as their facing area — length times thickness — grows.

The inductive part is less intuitive. The aggressor’s current creates a magnetic field that links the victim loop, inducing a series voltage along the victim in the same direction as the aggressor’s current flow. For a victim terminated at both ends, that series voltage drives current that splits into both ends.

The key asymmetry is the direction. Capacitive coupling injects current at every point along the victim, and each injected current splits into two waves travelling away from that point — toward the near end and the far end equally. Inductive coupling induces a voltage whose resulting current flows in a loop, which means it adds at the near end and subtracts at the far end. The two mechanisms therefore reinforce at the near end; at the far end they subtract, and how completely they cancel depends on the medium.

Near end and far end are different problems

The near-end crosstalk — the noise appearing at the end of the victim nearest the aggressor’s driver — accumulates as the aggressor’s wave travels forward. Each increment of length contributes an amount, and the contributions arriving back at the near end are delayed by the round trip to their point of origin. The total saturates once the coupling length exceeds half the rise time’s electrical length: contributions from further away arrive too late to overlap with the edge, and the noise reaches a plateau.

The plateau value is what datasheets and layout guides quote. In the weak-coupling limit the near-end coefficient saturates at roughly $(C_m/C + L_m/L)/4$, because at the near end the capacitive and inductive contributions add. Stripline does not cancel that sum: buried between two planes, its field is confined to a thinner volume, so both coefficients are smaller than a microstrip pair’s at the same spacing, but the near end still adds them. The important property is that the coefficient does not keep growing with length. A long parallel run is not much worse than a moderate one for near-end noise.

Far-end crosstalk behaves oppositely. It accumulates in proportion to coupling length because every increment’s contribution travels forward with the aggressor’s wave and arrives at the far end at roughly the same time. It is therefore the dominant concern for long parallel runs on an outer layer, and the reason “minimise parallel length” appears in every high-speed layout guide.

Far-end crosstalk depends on the medium in a stronger way. In a homogeneous dielectric — stripline — the capacitive and inductive coupling coefficients are equal, so the forward-travelling contributions from every point along the run arrive together and cancel; ideal stripline has no far-end crosstalk at all. On microstrip the dielectric is only below the trace, the two coefficients are unequal, cancellation is incomplete, and whatever difference remains accumulates with length. The far-end component is therefore much smaller on an inner layer, which is the practical argument for routing long parallel buses there: not that coupling is eliminated, but that the one component which accumulates with the run is.

Why the coupled length matters more than the spacing

Coupling falls steeply once the gap is large compared with the dielectric height: in the weak-coupling regime the coefficient goes roughly as $(h/s)^2$, so doubling the gap there reduces the coupled amplitude by about a factor of four. Close in — a gap comparable to the height — the curve is much flatter, and the first increment of spacing buys more than the second.

But the coupled amplitude at the far end is the coupling coefficient multiplied by the length over which it acts, normalised to the edge length. Doubling the length while keeping the spacing doubles the far-end noise. That is why a design rule stated as “keep parallel runs under N millimetres” is more effective than one stated as “keep spacing above M millimetres”: the length is a linear factor and the spacing is roughly quadratic, but the achievable range of lengths in a real layout is much wider than the achievable range of spacings.

Spacing rules are clearer in dielectric heights than in trace widths. A gap of three trace widths is about four heights on a 0.1 mm dielectric with 0.15 mm traces, and by the $(h/s)^2$ scaling that puts the coupling at a small fraction of what it is at a gap comparable to the height. A gap of one trace width, by contrast, is barely more than the height, which is where the curve is still flat and extra spacing buys little.

The role of the reference plane

The return current argument in Return Current Paths and Why Loop Area Sets Trace Inductance explains why the plane matters here. Each trace’s field is confined between itself and the plane; when two traces are close, their fields overlap in the region between them, and that overlap is the coupling. A thinner dielectric confines the field more tightly and reduces the overlap at a given spacing, which is why dense boards with thin dielectrics can route traces closer together than thick boards with the same crosstalk budget.

The converse is also true and is often missed. A trace with no plane beneath it — a route that crosses a plane split, or one on a layer whose reference is a power plane with a large void — has a much wider field and couples strongly to everything nearby. Fixing the reference fixes the coupling as a side effect.

Two traces on adjacent layers couple vertically as well as horizontally. Broadside coupling between a trace on layer 3 and a trace on layer 5, with a plane between them, is small because the plane shields. Without an intervening plane, or with a plane that has a void, broadside coupling can exceed the edge-to-edge coupling on the same layer, because the traces overlap for their entire length rather than only along their edges. The standard rule is to route orthogonal — horizontal on one layer, vertical on the next — so that any overlap is short.

Guard traces: when they help and when they hurt

A guard trace is a conductor routed between the aggressor and the victim, tied to ground at intervals. The intent is to intercept the coupling field and shunt it to the reference.

Its effectiveness depends entirely on how it is connected. A guard trace stitched to the plane with vias at both ends and at intervals shorter than a tenth of the edge’s electrical length acts as a wall: the field terminates on it and the coupling falls substantially. A guard trace connected only at one end is a resonant stub — it absorbs coupling at some frequencies, radiates at others, and can make the victim’s noise worse than it was without the guard.

The stitching interval is the parameter that matters. Vias spaced further apart than the edge length let the guard float between them, and the floating section re-radiates into both traces. Placing vias at both ends of the guard, and at intervals along it, is the only version that reliably works. Even then, a guard trace costs routing area and adds capacitance to both neighbours, which lowers their impedance and produces its own reflections.

For most designs the cheaper fix is spacing plus a continuous reference plane. Guard traces are reserved for cases where the coupling budget is tight and the geometry is fixed — a connector breakout, or a region where a clock must run beside a sensitive analog input.

Quantifying the budget

Crosstalk is usually specified as a fraction of the aggressor’s swing. A receiver with a 200 mV noise margin on a 1.8 V swing tolerates about 11% of the swing as coupled noise, but that budget must also cover reflections, power supply noise, and jitter, so a practical allocation for crosstalk is a few percent.

PS_PER_MM = 6.5

def near_end(k, coupled_len_mm, rise_ps):
    """NEXT saturates once the run exceeds half an edge length."""
    edge_mm = rise_ps / PS_PER_MM
    return k * min(coupled_len_mm, edge_mm / 2) / (edge_mm / 2)

def far_end(k, coupled_len_mm, rise_ps):
    """FEXT accumulates over the whole run, scaled by edge length."""
    edge_mm = rise_ps / PS_PER_MM
    return k * coupled_len_mm / edge_mm

# 20 mm parallel run; k is an illustrative saturated coupling coefficient
for rise in (200, 500, 1000):
    print(f"rise {rise:4d} ps: NEXT {near_end(0.02, 20, rise):.3f}"
          f"  FEXT {far_end(0.02, 20, rise):.3f}")

The k is an illustrative coefficient, and the point of the calculation is the shape rather than the constant. Near-end noise stops growing once the run exceeds about half an edge length, because contributions from further along arrive too late to overlap the edge. Far-end noise keeps growing with the run, but the same run produces less of it when the edge is slow, because the coupled contributions from different points along the run no longer arrive together.

Differential pairs and crosstalk

A differential pair is more immune to crosstalk than two single-ended traces, but the immunity is not absolute. Coupled noise that arrives equally on both conductors is common-mode and is rejected by the receiver’s differential input. Coupled noise that arrives unequally — because one conductor is closer to the aggressor, or because the pair is asymmetric — becomes differential and is amplified by the same factor as the signal.

The practical consequence is that a pair routed beside a noisy net should be oriented so that the aggressor is equidistant from both conductors, which usually means routing the pair broadside to the aggressor rather than parallel to one of its members. When that is impossible, the pair’s differential impedance and its common-mode rejection both degrade, which is the asymmetry argument in Differential Pairs and the Cost of Asymmetry.

Crosstalk also couples into the reference planes themselves, where it appears as ground bounce. The plane’s own impedance is not zero, so injected current produces a voltage difference between two points on the same plane, and every single-ended signal referenced to that plane sees it as noise. That path is covered in Power Delivery Network Impedance and the Target It Must Hold, where the plane impedance is the quantity being designed rather than an accident.