Signal Integrity on the Interface

Generic engineering material, applicable to any module from any supplier.

For more on organizing work around technical projects, read learn more.

A camera interface carries fast digital signals over a cable and a board, and at those speeds the wiring stops being wiring — it is a transmission line, and it behaves like one whether or not it was designed as one.

What the failures look like

Corrupted frames, whole or partial, appearing intermittently.

A link that works at one resolution and fails at a higher one, since the data rate rose and the margin did not.

A useful external technical reference is Springer.

Behaviour that changes when the cable is moved, which is the clearest signal of a marginal link.

And a system that works on the bench and fails in the enclosure, where the routing is different and other things are running.

None of these looks like a signal integrity problem. They look like a faulty module, which is why the module gets replaced first and the fault persists.

The rules

Match the impedance. Differential pairs have a specified impedance, and the board stack-up must be designed to it — not assumed.

Keep the pairs together and matched in length, since a difference between the two conductors of a pair becomes a timing error.

Keep the reference plane continuous underneath. A gap in the ground under a high-speed trace is a discontinuity, and it is the most common board-level cause.

Avoid stubs and vias where possible, and where a via is unavoidable, keep it short.

And respect the cable length. The qualified lengths exist because somebody tested them, and exceeding them is not a small extrapolation.

Why length matters more than it seems

Loss rises with frequency and with distance, so a longer cable attenuates the fast edges more than the slow ones — which closes the eye at the receiver.

And reflections from any impedance mismatch travel back and forth, arriving at times that depend on length: a length that works may be one where the reflection arrives harmlessly, and a slightly different one may not.

Which is why "it worked at 15 cm so 30 cm will be fine" is not a safe inference, and why suppliers qualify specific lengths rather than a maximum.

Diagnosing it

Reduce the data rate. Drop the resolution or the frame rate — if the errors stop, the problem is margin rather than a fault.

Shorten the cable. Test with the shortest available, which isolates the cable from the board.

Swap the cable for a known-good one, which distinguishes a damaged cable from a marginal design.

And check the frame error rate over time rather than watching, since an intermittent fault needs statistics rather than an impression.

Where the errors correlate with machine state, the cause is coupling rather than the link itself — a different problem with a different remedy.

What the supplier's reference design is for

It is a worked answer to everything on this page, and it is frequently treated as an example rather than as a specification.

The trace widths, the spacing, the stack-up and the component placement were arrived at by somebody who tested the result.

Departing from them is permitted and costs verification. Where a layout must differ — and it usually must — the departures should be deliberate and few, rather than the result of routing convenience.

And where a reference design does not exist for your interface, that is a selection finding rather than a detail: it moves work from the supplier to you.

Termination and the receiver

The far end of a transmission line needs terminating, and on most camera interfaces the receiver does it internally.

Which means the configuration matters: a receiver with termination disabled, or with the wrong mode selected, produces exactly the symptoms above while the hardware is correct.

Check the processor's camera interface configuration before suspecting the board — it is a register setting and it is faster to verify than anything physical.

Margin, and why it should be measured

A link either works or it does not, which conceals how close it came.

Which means a design with no margin passes every test and fails in production, as component tolerances and cable variation move it across the line.

Measure the margin where the tools exist: an eye diagram at the receiver, or an error rate at a deliberately raised data rate.

Where they do not, the practical substitute is testing at a higher resolution than the product will use — if it works there, the production configuration has room.

The board that was fine last revision

A layout change made for an unrelated reason — a component moved, a plane split for a new supply, a connector relocated — can break a link that was working.

Which is why the high-speed routing should be marked as such on the board, and why a change near it should trigger a check rather than an assumption.

A note in the design file costs nothing and prevents a revision that fails for a reason nobody connects to the change that caused it.

Testing at the extremes

A link with adequate margin at 20 degrees may not have it at 60.

Loss rises with temperature and component values shift, which makes a thermal chamber test of the link worth doing on anything going into a warm enclosure.

And at the cold end, where a flexible cable stiffens and its electrical properties change slightly.

Run the frame error count at both extremes rather than assuming the bench result holds — which is the same argument as everywhere else in this section.

In one line

Use a qualified cable length and follow the reference layout, because the failures look like a broken module.

The short version

• At camera interface speeds the wiring is a transmission line whether or not it was designed as one
• Failures look like a faulty module: corrupted frames, working at one resolution and not a higher one, changing when the cable moves
• Rules: matched impedance, matched pair lengths, a continuous reference plane, minimal stubs and vias, and the qualified cable length
• Loss rises with frequency and distance, and reflections arrive at times that depend on length — so a working length does not extrapolate
• Diagnose by reducing the data rate, shortening the cable, and swapping for a known-good one
• Errors correlated with machine state are coupling rather than the link, which is a different problem entirely