Skip to main content
Part 2 · Medium: the trace is real

15 · Crosstalk

Status: 🟡 Runs, but don't trust the numbers yet. The two-lane extraction runs, but its coupling doesn't behave physically as spacing changes. That makes this lesson about checking a model as much as about crosstalk.

Goal: measure how much a switching trace disturbs its quiet neighbour, and sanity-check the result.

The idea​

Coupled traces share mutual inductance (Lm) and mutual capacitance (Cm). When the aggressor switches, the victim sees two kinds of noise:

  • NEXT (near-end crosstalk) at the driver end. It scales with (Cm/C + Lm/L).
  • FEXT (far-end crosstalk) at the receiver. It scales with (Cm/C − Lm/L) × length. In microstrip it's usually negative, because inductive coupling dominates.

Both should fall as the spacing grows.

Build​

The aggressor (IO1) and the victim (IO2) run straight and parallel for 20 mm. The chips' I/O pad pitch sets the spacing. The victim is driven but held at 0:

15-crosstalk.circuit.tsx
import { LineBoard } from "./teaching-board"

// Aggressor (IO1) and victim (IO2) run straight and parallel for 20 mm.
// `spacing` is centre to centre in mm; the chips' I/O pads use the same pitch.
export const Crosstalk = ({ spacing = 0.32 }: { spacing?: number }) => (
  <LineBoard name="XTALK" drive="LVCMOS33_24MA" victimSpacing={spacing} />
)

export default () => <Crosstalk spacing={0.32} />

LineBoard adds the second lane, a quiet stimulus (bits: "0"), and VIC_NEAR/VIC_FAR probes when victimSpacing is set. The spacing variants are 15-crosstalk-032, -064 and -128.

What you see​

24 mA aggressor, 2 ns per bit, peak victim noise:

Spacing (centre to centre)Lm/LCm/CNEXT peakFEXT peak
0.32 mm (gap = 1 trace width)+0.0660.14571 mV75 mV
0.64 mm (gap = 3 trace widths)−0.0190.04651 mV107 mV
1.28 mm (gap = 7 trace widths)+0.0170.02013 mV27 mV

Check by hand, and why this fails the check​

  • Cm/C falls about 3× from 0.32 mm to 0.64 mm. That's sensible.
  • Lm/L goes negative at 0.64 mm. Two parallel traces over a solid plane always have positive mutual inductance, so this is not physical.
  • FEXT grows from 0.32 mm to 0.64 mm, the opposite of what it should do, before falling at 1.28 mm.
  • From 0.32 mm to 1.28 mm overall, NEXT falls about 5× and FEXT about 3×. The end-to-end trend is right; the middle point is the warning sign.

The cause is the model, not the board. straightened-stages puts both traces over one 2 mm-wide return bar, not a whole ground plane. At 0.64 mm the traces sit near the bar's edges, and the bar's own current crowding produces the odd mutual terms. A real plane is tens of mm wide.

This is the habit the lesson teaches: when a trend goes the wrong way, suspect the model before the physics.

What would fix it​

  • A return bar much wider than the trace spacing. Try straightenedStageProfiles with a custom return.width.
  • reduction: "routed-copper", which meshes the real pour, once it converges (see lesson 20).

Try this​

Add a straightenedStageProfiles entry with a 10 mm return width, then check whether Lm/L turns positive and FEXT falls with spacing.