Crosstalk Analysis — Theory and 3D FEM Evaluation of NEXT/FEXT

Put simply, between adjacent wires there exist parasitic capacitance and mutual inductance. When a signal flows through one wire (the aggressor), its electric and magnetic fields couple to the adjacent wire (the victim), inducing an unintended noise voltage. This is crosstalk.

That's a good analogy. Exactly like that, the thinner the wall (dielectric), the easier the sound (signal) leaks. On a PCB, the narrower the wire spacing and the longer the parallel run, the greater the crosstalk becomes. And there are two types of noise that leak through.

• NEXT (Near-End Crosstalk): Noise observed on the side closer to the signal source. Also called backward crosstalk.
• FEXT (Far-End Crosstalk): Noise observed on the far-end side where the signal arrives. Also called forward crosstalk.

Because their physical mechanisms are completely different. For NEXT, the capacitive coupling and inductive coupling add up, so it always occurs. On the other hand, FEXT is determined by the difference between the two, so if capacitive and inductive coupling are perfectly balanced, FEXT becomes zero. This balance is easier to achieve in a stripline structure, but in a microstrip structure, the dielectric is only on one side, so the balance is broken and FEXT becomes larger.

Capacitive coupling is a phenomenon where electric fields couple through the parasitic capacitance $C_m$ (mutual capacitance) existing between two conductors. A noise current proportional to the voltage change $dV/dt$ in the aggressor wire flows into the victim via $C_m$.

On the other hand, inductive coupling is magnetic field coupling through the mutual inductance $L_m$ between wires. The current change $dI/dt$ flowing in the aggressor induces an electromotive force in the victim.

Exactly. The faster the signal edge rate (shorter rise time), the larger $dV/dt$ and $dI/dt$ become, worsening crosstalk. For example, with PCIe Gen5 at 32 GT/s, crosstalk of about -20 dB can occur even over a 1mm parallel section. It wasn't that critical in the DDR4 era, but now management is needed at the level of each individual wire.

Under the weak coupling approximation (when the coupling amount is sufficiently small) and assuming uniform coupled lines, the NEXT and FEXT coefficients are expressed as follows.

NEXT (Near-End Crosstalk) coefficient:

Here, $C_m$ is the mutual capacitance per unit length [F/m], $L_m$ is the mutual inductance per unit length [H/m], and $T_d$ is the propagation delay of the coupled section [s].

FEXT (Far-End Crosstalk) coefficient:

Here, $\ell$ is the length of the coupled section [m].

Sharp observation. When $C_m = L_m$, meaning capacitive and inductive coupling are perfectly balanced, FEXT becomes zero. In a stripline (inner layer trace) surrounded by a uniform dielectric, this condition approximately holds, so FEXT becomes very small. This is one reason why it's said "route high-speed buses on inner layers."

Conversely, a microstrip (outer layer trace) has an asymmetric structure with air above and dielectric below, so $C_m \neq L_m$ and FEXT becomes larger. In practice, -30 dB is often used as a threshold, and if exceeded, design changes are considered.

Yes. However, NEXT has another important characteristic. The amplitude of NEXT hardly depends on the length of the coupled section. Beyond a certain length (the saturation length $\ell_{\text{sat}}$), NEXT saturates to a constant value. The saturation length can be estimated by:

Here, $t_r$ is the signal rise time, $v_p$ is the propagation velocity. For example, for a signal with a 50 ps edge rate on an FR-4 substrate ($v_p \approx 1.5 \times 10^8$ m/s), $\ell_{\text{sat}} \approx 3.75$ mm. Parallel sections longer than this will not increase NEXT no matter how much they are extended.

There are several definitions for the coupling coefficient, but the two main ones are as follows.

Capacitive coupling coefficient:

Inductive coupling coefficient:

$C_{11}$ is the self-capacitance of wire 1, $L_{11}$ is the self-inductance. The coupling coefficient ranges from 0 to 1, where $k = 0$ means completely independent and $k = 1$ means completely coupled. In practice, $k < 0.05$ (below -26 dB) is often targeted.

Exactly. Coupled wires are modeled as coupled transmission lines. Solving for the TEM/quasi-TEM modes of two coupled lines yields two eigenmodes: the even mode and the odd mode.

The differential impedance is $Z_{\text{diff}} = 2 Z_{\text{odd}}$, and the single-ended characteristic impedance is close to the geometric mean $Z_0 = \sqrt{Z_{\text{even}} \cdot Z_{\text{odd}}}$. The standard approach is to solve the electromagnetic field in the cross-section using 3D FEM, obtain the electric field distributions for the even and odd modes, and extract the PUL (Per-Unit-Length) parameter matrices [$C$], [$L$] from them.

Let me show some specific numbers.

As frequency increases, self-attenuation (insertion loss) becomes larger due to skin effect and dielectric loss, but at the same time, the allowable margin for crosstalk also decreases. For PCIe Gen6, with PAM4 modulation, the level spacing becomes 1/3, so the impact of crosstalk becomes about 3 times more severe than NRZ.

The most basic method is to extract PUL (Per-Unit-Length) parameters using 2D cross-section FEM analysis. Model the cross-sectional shape of the wires (trace width, thickness, spacing, dielectric stack-up) and solve for the electrostatic and magnetostatic fields separately.

Electrostatic analysis: Apply potential $V_1 = 1\text{V}$ to wire 1, $V_2 = 0$ to wire 2, and solve Laplace's equation:

From the solution, calculate the charge $Q_i$ on each wire and derive the capacitance matrix:

The inductance can be calculated using the capacitance matrix $[C_0]$ with all dielectrics replaced by vacuum:

This relationship is based on the assumption of TEM mode propagation.

2D PUL extraction assumes "an infinitely uniform cross-section continues," so 3D analysis becomes essential in the following cases:

• Via transition sections: Where wires move between layers, the cross-section is not uniform.
• Connector pin arrays: Crosstalk between adjacent pins.
• Branching/bending sections: Where wires bend, impedance discontinuities occur.
• BGA fan-out: Radial trace patterns.
• Frequency > 10 GHz: The quasi-TEM assumption begins to break down.

In 3D FEM, Maxwell's equations are formulated using the vector potential $\mathbf{A}$. In the frequency domain:

Here, $k_0 = \omega \sqrt{\mu_0 \varepsilon_0}$ is the free-space wavenumber. This equation is discretized using edge elements (Nedelec elements). The reason for using edge elements is that they automatically guarantee the continuity of the tangential electric field components and eliminate spurious (non-physical, false) solutions.

After discretization, the system equation becomes:

$[S]$ is the curl-curl matrix, $[T]$ is the mass matrix. Solve this using direct or iterative methods to obtain the electromagnetic field distribution, and then extract S-parameters from it.

For a 4-port coupled wire model (ports 1,2: near end, ports 3,4: far end), crosstalk is evaluated using the following S-parameters.

The dB expression for crosstalk is:

For example, if $|S_{31}| = 0.05$, then $\text{NEXT} = -26$ dB. Generally, NEXT < -25 dB and FEXT < -30 dB are design targets for high-speed SI.

There are two main methods to convert from S-parameters (frequency domain) to the time domain.

• IFFT-based time domain conversion: Apply inverse FFT to S-parameters to obtain the impulse response, then convolve it with the input waveform. This method is used in SPICE W-elements, etc.
• Rational function approximation (Vector Fitting): Approximate S-parameters with a rational function $S(s) \ldots$