Differential Signalling

Chapter 3.1 distinguished J from K by asking which conductor is higher than the other. That phrasing was doing quiet work, and this chapter unpacks it.

It means the receiver's decision is a function of the difference between the conductors and not of either conductor's level against ground. That is a deliberate and consequential choice, and the usual one-line justification for it — noise immunity — is simultaneously correct and the source of a genuinely damaging misconception. Differential signalling does not remove noise. What it does is narrower, conditional, and worth understanding precisely, because the conditions it depends on are exactly the ones a board designer can violate.

1. The Problem With Measuring Against Ground

Start with the alternative, because the case for differential is a case against something.

A single-ended interface represents information as a conductor's voltage measured against a shared ground reference. The receiver compares that conductor against a threshold and decides. Chapter 1.3's RS-232 works this way, and so does ordinary logic inside a chip.

That arrangement carries two assumptions that are easy to satisfy on a circuit board and hard to satisfy across a cable between separately powered boxes.

It assumes the two ends agree about ground. They frequently do not. Currents flowing in the ground path, different supply arrangements, and simple distance mean the receiver's idea of zero volts can differ from the transmitter's. A signal referenced to the transmitter's ground is then being measured against a different zero — and the whole measurement shifts.

It assumes interference does not move the conductor relative to that reference. Any coupled disturbance adds directly to the quantity being measured, because the measurement is the conductor's level.

Both assumptions degrade with cable length and with the electrical hostility of the environment. An interface designed to run several metres to a device with its own power supply cannot rely on either.

2. Measuring the Difference Instead

A differential pair carries the information in the relationship between two conductors, with the transmitter driving them to opposite states and the receiver deciding on their difference.

The mechanism is subtraction. If a disturbance adds the same amount to both conductors, it contributes to both terms of the difference and cancels. A ground offset between the two ends does the same thing: it shifts both conductors together as seen by the receiver, and a measurement that never looks at either conductor against ground is insensitive to it.

Interference that affects both conductors equally is called common-mode; the driven signal, which moves them oppositely, is the differential-mode component. The receiver is built to respond to the second and ignore the first.

3. The Same Signal, With and Without a Disturbance

4. What It Does Not Buy

Now the part that the one-line justification omits, and that separates an engineer who can use this idea from one who has only heard it.

Rejection is not cancellation, and it is finite. A real differential receiver responds slightly to common-mode input; its ability to ignore it is a bounded quantity, not an absolute. Push the common-mode level far enough — beyond the range the receiver's inputs are specified to operate over — and the receiver stops working correctly regardless of how clean the difference is. A large enough ground offset is not rejected; it takes the receiver outside its operating conditions.

The benefit is conditional on the disturbance arriving equally. That is a physical routing property, not something the signalling scheme guarantees on its own. It holds when the two conductors run close together, over the same path, with matched lengths and similar surroundings, so that whatever couples into one couples into the other at nearly the same instant and amplitude. Separate the conductors, route them differently, mismatch their lengths, or let one pass near an aggressor the other does not, and a disturbance that would have been common-mode becomes partly differential — arriving inside the question rather than outside it.

It does nothing about disturbances that are already differential. Crosstalk that couples predominantly into one conductor is differential by construction and is not rejected.

It does not address reflections. Impedance discontinuities produce reflections on the pair itself, which are part of the differential signal and therefore not rejectable by any amount of common-mode rejection. That is a termination and impedance-control problem, and Chapter 3.8 is where it belongs.

And it costs conductors. Two per signal rather than one, plus the routing discipline above. That cost is exactly why the interface uses one differential pair carrying data in both directions at different times rather than several.

The honest summary: differential signalling converts a large class of real-world disturbance from fatal to irrelevant, provided the pair is built and routed so that the class actually applies. It is a powerful technique with stated preconditions, and the preconditions are a physical-design responsibility rather than a property of the protocol.

5. What Follows for the Digital Side

Two consequences reach RTL and verification engineers directly, and both are about what your model can and cannot claim.

Your digital model does not contain any of this. Chapter 3.1 §5's decoder takes already-thresholded receiver outputs. Nothing in it represents a margin, a common-mode level or a rejection ratio, and no simulation of it can tell you whether the analog layer is working. A passing digital regression is evidence about decode logic and evidence about nothing else. Stating that boundary explicitly in a verification plan is more useful than any number of additional digital tests.

The failure signature is distinctive, and it is not a logic signature. Analog-layer degradation produces errors that are rate-dependent, cable-dependent, temperature-dependent, board-dependent and statistical — a small fraction of transfers failing, more at higher rates or longer cables, irreproducible in simulation because simulation has no cable. Digital logic errors are deterministic: the same input sequence produces the same wrong answer every time. When a symptom is statistical and correlates with physical variables, adding digital tests is the wrong response, and the right instruments are at the electrical layer rather than in a waveform viewer.

That distinction — is this failure deterministic or statistical? — is one of the most useful first questions available when a physical interface misbehaves, and it costs nothing to ask.

6. Common Misconceptions

7. Reason It Through

A product works on a short cable and accumulates errors on a long one. The error rate rises with cable length and with ambient electrical activity. Two units behave differently on the same cable. Digital simulations are clean.

Deterministic or statistical? Statistical, and correlated with physical variables. That alone moves the investigation out of the digital domain, because no logic fault varies with cable length or with what else is switched on nearby.

Why do digital simulations tell you nothing here? Because they contain no cable, no coupling and no margin. They can confirm the decode logic is correct, and the evidence says the decode logic is not the problem.

Why might two units differ? Because the margin available is a physical property of each assembly — connector seating, routing, component tolerance, ground arrangement. Identical designs have non-identical margins, so a marginal design shows up as unit-to-unit variation, which is itself a signature worth recognising.

Which of §4's mechanisms are candidates? Rising error rate with length suggests accumulated attenuation and degradation reducing the difference the receiver must resolve. Correlation with ambient activity suggests coupling — and if it is being coupled unequally, because of routing or a mismatch, it is differential and not rejected. A ground offset large enough to push the receiver outside its operating range is a third candidate and is testable by changing the grounding arrangement.

What would you actually do? Measure at the electrical layer, because that is the only place the quantity in question exists. And stop adding digital tests: a hundred more would pass, and each one would be evidence about something that has already been shown not to be broken.

The general lesson. Match the instrument to the abstraction the failure lives in. A statistical, physically-correlated symptom lives below the digital model, and the digital model's silence about it is a property of the model rather than evidence of health.

8. Understanding Check

9. Summary

A single-ended interface measures a conductor against a shared ground reference, which assumes both ends agree about ground and that nothing has moved the conductor relative to it. Across a cable to separately powered equipment, neither assumption holds well.

A differential pair carries the information in the relationship between two conductors. Because the receiver's decision is the difference, a disturbance arriving equally on both — interference coupled into a closely routed pair, or a ground offset between the two ends — contributes to both terms and cancels. Noise is not removed; it is made irrelevant to the question being asked.

The benefit has stated preconditions, and they are physical-design responsibilities rather than protocol guarantees. Rejection is finite, and a large enough common-mode excursion takes the receiver outside its operating range. The benefit depends on the disturbance arriving equally, which requires the conductors to be closely coupled, similarly routed and length-matched. Disturbances that are already differential — crosstalk into one conductor, reflections from impedance discontinuities — are inside the measurement and are not rejected.

For the digital side, two things follow. A digital model contains no margin and no common-mode quantity, so a passing regression is evidence about decode logic and about nothing else. And the failure signatures differ: analog degradation is statistical and physically correlated, logic errors are deterministic. Asking which one you have is cheap, and it selects the instrument.

10. What Comes Next

This chapter and the last have both treated the pair as symmetric — two conductors, a difference between them, a decision. That symmetry is real at the level of the measurement, and it is misleading about everything else.

Chapter 3.3 takes D+ on its own and shows that the pair is not symmetric in its roles. One conductor carries responsibilities the other does not, the idle convention is expressed in terms of specific conductors, and a device announces its own presence by acting on a particular one. Understanding which conductor does what — and then, in Chapter 3.4, discovering that the answer changes with speed — is what finally dismantles the idea that J and K are fixed voltage patterns.

Browse the full path on the USB tutorials index.