Crosstalk Model
How Coupled-Line Theory Predicts Crosstalk
A crosstalk model treats two parallel conductors as a coupled-line system whose interaction is captured by a 2x2 capacitance matrix and a 2x2 inductance matrix. The off-diagonal entries, mutual capacitance Cm and mutual inductance Lm, set how much of the aggressor signal appears on the victim. Capacitive coupling injects a displacement current proportional to dV/dt, and inductive coupling injects a series voltage proportional to dI/dt. Both effects act simultaneously, and their vector sum and difference define the two distinct noise events seen at each end of the victim trace.
The defining behavior is that the two coupling mechanisms add at the near end and subtract at the far end. Near-end crosstalk reaches a saturated backward-coupling coefficient once the coupled section exceeds roughly half the signal rise length, so lengthening a closely routed bus does not keep increasing NEXT. Far-end crosstalk, by contrast, grows linearly with coupled length and with edge rate, which is why a long, fast bus can show a large FEXT spike even when the spacing looks generous. In a homogeneous medium such as stripline, where the dielectric is uniform above and below, CmZ0 and Lm/Z0 nearly cancel and FEXT collapses toward zero.
Practical model accuracy depends on extracting Cm and Lm correctly. Closed-form approximations work for simple edge-coupled microstrip, but field solvers are used when the geometry includes reference-plane gaps, vias, or non-uniform spacing. The model is then validated against time-domain or vector network analyzer measurements; a verified extraction lets the designer trade spacing, dielectric height, and guard traces against an isolation target without repeated board spins.
Governing Crosstalk Equations
KNEXT ≈ ¼ × (CmZ0 + Lm/Z0)
Far-end (forward) crosstalk voltage:
VFEXT ≈ ½ × (CmZ0 − Lm/Z0) × (L / Tr) × (dV/dt)
Isolation in decibels:
Isolation = 20 × log10(Vcoupled / Vaggressor) dB
Where Cm = mutual capacitance per unit length, Lm = mutual inductance per unit length, Z0 = line characteristic impedance, L = coupled length, Tr = aggressor rise time. Example: KNEXT ≈ 0.02 → isolation ≈ −34 dB.
Spacing, Coupling, and Isolation
| Edge spacing (S/h) | NEXT @ 1 GHz | FEXT @ 1 GHz | Coupling level | Typical mitigation |
|---|---|---|---|---|
| 1× (touching) | −15 dB | −20 dB | Heavy | Generally unacceptable |
| 2× | −25 dB | −30 dB | Moderate | Add guard trace |
| 3× (3W rule) | −35 dB | −40 dB | Low | Standard design rule |
| 5× | −45 dB | −50 dB | Very low | Sensitive RF / clock nets |
| Grounded guard between | −50 dB | −55 dB | Minimal | Best practical isolation |
Frequently Asked Questions
What is the difference between the NEXT and FEXT terms in a crosstalk model?
NEXT is the backward-coupled noise that returns toward the source; it is the sum (CmZ0 + Lm/Z0) and saturates to a constant once the coupled line is electrically long. FEXT is the forward-coupled noise at the far termination; it is the difference (CmZ0 − Lm/Z0) and grows with coupled length and edge rate. In homogeneous stripline the difference term nearly cancels, so FEXT approaches zero, while inhomogeneous microstrip leaves a residual far-end spike.
How do you convert modeled crosstalk voltage into an isolation figure in dB?
Take the ratio of coupled victim voltage to aggressor swing and apply 20×log10. A modeled NEXT coefficient of 0.05 equals −26 dB; the 3W spacing rule typically gives about 0.02, or roughly −34 dB. For connectors and cable assemblies the lumped coupled-line form is replaced by a shield transfer impedance model, since aperture and braid leakage dominate over per-unit-length L and C.
Why does a crosstalk model depend so strongly on the aggressor rise time?
Capacitive coupling scales with dV/dt and inductive coupling with dI/dt, so faster edges carry more spectral energy into the victim. In the FEXT expression the noise scales as coupled length divided by 4 times the rise time, meaning a 10× faster edge yields about 10× more far-end noise. That sensitivity is why edge-rate control, series termination, and slew-limited drivers remain effective mitigations even when the routing geometry is already fixed.