Crosstalk Analysis — NEXT/FEXT Theory and 3D FEM Evaluation
Theory: two couplings, electric and magnetic
Overview
On our PCB, noise appears on a trace that isn't carrying anything. I suspect the neighbouring trace, but…
That's crosstalk. When the voltage on the neighbouring aggressor changes, current flows through the mutual capacitance; when its current changes, the mutual inductance induces a voltage. On the victim line these two couplings add at one end and subtract at the other — the difference between near-end (NEXT) and far-end (FEXT) crosstalk.
Near-end and far-end formulas
$C$ and $L$ are self-capacitance and self-inductance per unit length, $C_m$ and $L_m$ the mutual values, $\ell$ the coupled length, $t_r$ the rise time and $v$ the propagation velocity. These approximations assume weak coupling.
Homogeneous dielectric (stripline)
If the traces are completely surrounded by one dielectric, the capacitive and inductive ratios are equal and far-end crosstalk vanishes in principle. In microstrip, with air above and substrate below, part of the field runs through air, so $C_m/C < L_m/L$ and negative noise appears at the far end.
Saturation of near-end crosstalk
Near-end noise stops growing once the coupled section's round-trip time exceeds the rise time. With a 100 ps edge on FR-4-like material (relative permittivity 4, 15 cm/ns), the saturation length is about 0.75 cm. On most board traces, near-end crosstalk can be assumed saturated.
A problem that began with crossed telephone lines
The word crosstalk comes from telephone lines, where faint conversations from other lines could be overheard. By the late nineteenth century, parallel telephone wires were already being transposed — swapped or twisted at intervals — to cancel induction from neighbouring lines. Today's twisted-pair cables and differential routing on high-speed boards extend the same idea of cancelling coupling. What has changed is frequency: from voice at a few kilohertz to digital signals at tens of gigahertz, with rise times shorter by orders of magnitude.
Worked example: spacing effect from a 2D field solution
Method
Two traces were placed in a stripline between two ground planes (homogeneous dielectric, relative permittivity 4, zero trace thickness) and the cross-sectional Laplace equation was solved by finite differences. With 1 V on one trace and 0 V on the other and both grounds, the charge on each trace gives the self-capacitance $C$ and mutual capacitance $C_m$. Because the dielectric is homogeneous, impedance and crosstalk follow from capacitance alone.
Check against the exact solution first
For a single stripline with width-to-spacing ratio $w/b = 0.5$, the exact elliptic-integral solution is $Z_0 = 50.25$ Ω. The finite-difference result stayed around 51.7 Ω even on a finer grid — about 3% high. The field is singular at the edges of a zero-thickness strip and the grid cannot fully capture the charge there, so the values below carry an error of similar size.
Spacing and near-end crosstalk
| Spacing | Saturated NEXT | Noise for a 1 V signal |
|---|---|---|
| 0.5 × width | 7.9 % | 79 mV |
| 1 × width | 3.5 % | 35 mV |
| 2 × width | 0.72 % | 7 mV |
| 3 × width | 0.15 % | 1.5 mV |
Going from one to two widths of spacing cuts it to a fifth.
In stripline the fields are pulled into the ground planes above and below and cannot spread sideways, so coupling falls steeply with spacing. That's behind the rule of thumb of spacing at three trace widths (the 3W rule). But faster signals have smaller noise budgets, and a victim often hears several aggressors at once, so compare the total from all neighbours with the budget.
Analysis and design practice
Workflow
- From the stack-up and trace cross-section, compute the capacitance and inductance matrices with a 2D field solver.
- From coupled length and rise time, estimate whether NEXT is saturated and how large FEXT is.
- Confirm the noise waveform in circuit simulation with real driver waveforms and terminations.
- For non-uniform parts such as vias and connectors, obtain S-parameters from 3D electromagnetic analysis and insert them into the channel model.
Tips for 3D analysis
- Always include the return path. Traces crossing gaps or slots in the return plane see much stronger coupling.
- Port placement and reference (what counts as ground) drive the result. Place near- and far-end ports correctly at both ends of each trace.
- Trace thickness, edge rounding and surface roughness matter more at higher frequencies; a few percent difference from a zero-thickness 2D estimate is normal.
- When converting frequency-domain results to time waveforms, you need continuous data from near DC up to a sufficiently high frequency; missing bandwidth produces unphysical ringing.
“We widened the spacing but the noise didn't drop”
When crosstalk won't go away despite generous spacing, the cause is often not trace-to-trace coupling but vias, connectors or a shared return. When several signals return through the same ground via or the same connector ground pin, noise couples through the inductance of that common path (ground bounce). A 2D trace analysis cannot see this, so use 3D analysis or measurements to locate the coupling.
Common mistakes
Mistakes and fixes
| Mistake | Effect | Fix |
|---|---|---|
| Treating microstrip FEXT like stripline | Far-end noise in microstrip is missed | Compute both capacitive and inductive ratios |
| Evaluating one aggressor only | Total noise underestimated | Sum all aggressors on both sides |
| Omitting the return path | Coupling badly wrong | Include planes and vias |
| Thinking only in clock frequency, not rise time | Fast edges create noise even with slow clocks | Use the rise time for saturation length and FEXT |
| Not checking solver accuracy | A few percent of error goes unnoticed | Check first on a geometry with an exact solution |
I'd like to try changing trace dimensions and see how crosstalk changes.
With the coupled-line crosstalk simulator you can vary spacing and coupled length and watch near- and far-end noise change. Differential routing can be explored with the differential pair impedance calculator, and the effect of noise on signal quality with the eye diagram simulator. Background is covered in the articles on transmission lines, differential signalling and S-parameters.
Related Topics
Related Fields
Helpful
Details
Error