Differential Pair Routing: Rules That Actually Matter

Differential pairs are surrounded by more folklore than almost any other part of high speed layout. Some of the advice is essential, some of it is a convention that no longer applies, and the difference shows up when a design fails its eye diagram test after being routed to the rules. This article separates the rules that decide whether a link works from the ones that only look tidy.

What a Differential Pair Actually Does

A differential pair carries one signal on two conductors, with the second conductor carrying the inverse. The receiver looks at the difference between the two, and any disturbance that appears equally on both conductors – which is what a nearby noise source or a shifting ground reference produces – is subtracted out and does not affect the decision.

That property is called common mode rejection, and it is the reason differential signalling survives in electrically hostile environments and at high data rates. It is also why differential pairs can run at lower voltage swings, which reduces the energy per bit and the radiated field.

The two conductors interact. The signal propagates in two modes at once: the odd mode, where the two currents are equal and opposite, and the even mode, where they are equal and in the same direction. The odd mode is what carries the data, and its velocity and impedance depend on how strongly the two traces couple to each other. That single fact explains most of the design rules.

The Rules That Matter

Three properties decide whether a pair behaves as intended, and everything else is secondary.

Constant spacing. The differential impedance is set by the trace geometry and the coupling between the two traces. If the spacing changes along the route, the impedance changes with it, and each change is a reflection. Keeping the spacing constant is more important than keeping it small.

Matched length. If one trace of the pair is longer than the other, the two signals arrive at different times. Part of the differential signal converts into a common mode signal, which radiates and which the receiver rejects imperfectly. The tolerance is set by the data rate, and for most multi gigabit links the budget is a small fraction of a bit period.

A continuous reference. Both traces need a solid return path underneath them. A split plane, a change of reference layer or a missing ground via at a layer transition turns the pair into an antenna and destroys the impedance control that the spacing was designed to provide.

Everything else – how tightly the pair is coupled, whether the two traces are routed side by side or far apart, whether the gap is adjusted to reach a target impedance – is a design choice rather than a rule, and it should be made deliberately rather than by copying a number from an older project.

Impedance and Coupling

A differential pair is quoted with two impedance figures: the single ended impedance of each trace to ground, and the differential impedance between them. The two are related, and the relationship depends on how strongly the traces couple.

A tightly coupled pair has the traces close together, so most of the field is between them. The differential impedance falls as the gap shrinks, which means a designer who narrows the gap to fit through a bottleneck must also narrow the traces to keep the impedance on target. A loosely coupled pair has the traces far apart, so each behaves almost like a single ended line and the differential impedance is close to twice the single ended impedance.

Both are valid. Tightly coupled pairs are narrower overall and take less routing area, but they are more sensitive to any change in spacing and harder to terminate consistently. Loosely coupled pairs are more forgiving of spacing variation and closer to a simple two line model, at the cost of width.

The choice matters most where the pair has to be broken. Every pair has to pass through a via field, around a component or through a connector, and each of those places forces the spacing to change. A loosely coupled pair barely notices; a tightly coupled pair needs compensation, which is where the designer ends up widening the gap and thinning the traces in the same section.

Length Matching

The mismatch between the two traces of a pair is called intra pair skew, and it is the parameter that most often explains a link that fails at the far end of its specification. It is also the one that is easiest to get quietly wrong, because a pair can be matched in total length and still mismatched over most of its route.

The way to avoid that is to correct a mismatch where it occurs rather than at the end. If one trace takes a detour around a via, the compensating length should be added to the other trace immediately afterwards, so that the two traces stay within the tolerance for the whole route. Adding the correction at the far end leaves a long section where the pair is unbalanced, and the common mode signal generated in that section has already been radiated by the time the correction is applied.

Where serpentine compensation is used, its shape matters. A tight serpentine with a small pitch couples into itself and creates its own impedance variation, which can cost more than the skew it corrects. A looser, larger amplitude serpentine is preferable. The same reasoning applies to inter pair matching on a parallel bus, where the tolerance is usually much wider than the intra pair budget and does not require aggressive tuning. Our article on trace skew covers the budgets and the measurement side in more detail.

Layer Transitions and Vias

Every change of layer is the most likely place for a differential pair to go wrong, and the problems come from three sources at once.

The first is the reference plane. A pair routed over a ground plane on one layer and over a power plane on the next layer changes its return path, and unless a via connects those planes close to the signal vias, the return current has to find another route. The standard practice is to place a ground via adjacent to the pair at every transition, so the return path is short and defined.

The second is the via itself. A via has capacitance and inductance, and a via that is not part of the impedance model creates a discontinuity. Where the pair changes layer, both traces need their own via, placed symmetrically, with the same stub length. A pair where one trace changes layer and the other does not is a guaranteed common mode source.

The third is the stub. On a thick board the unused portion of a through via behaves as a resonant stub that removes energy from the pair at a specific frequency. Where the stub is significant, back drilling or a blind via removes it, and this is one of the cases where a more expensive via structure produces a measurable improvement.

Symmetry is the guiding principle: if one trace of the pair has a via, so should the other, positioned so that the geometry around the transition is a mirror image as far as the layer stack allows. Asymmetry is what converts differential signal into common mode, and common mode is what radiates and what the receiver cannot reject.

Common Mistakes

• Splitting the pair to route around an obstacle and rejoining it, which changes the coupling and the impedance in the split region.
• Correcting intra pair skew at the end of the route rather than where it occurred.
• Routing one trace on a different layer from the other for any distance.
• Passing over a plane split, or over a region where the reference changes without a stitching via.
• Using a via count that differs between the two traces.
• Narrowing the gap to squeeze past a component without adjusting the trace width, so the impedance drops.
• Tight serpentine compensation that couples into itself and adds more disturbance than the skew it removes.
• Terminating the pair with two separate single ended terminations that are not symmetric, which introduces a common mode reflection.

Most of these are visible in a layout review if the pair is highlighted and inspected as one object rather than two traces. That is the practical technique: route the pair as a unit, review it as a unit, and check the impedance profile along the route rather than only at a test coupon.

Verifying the Result

Two measurements confirm that a pair has been routed as designed. A TDR measurement on a test coupon or on the board itself shows the differential impedance along the route and reveals discontinuities at vias, connectors and places where the spacing changed. A time domain measurement of the two traces gives the intra pair skew directly, and it is the number that the specification actually limits.

At the system level, the consequence of a badly routed pair is jitter and an eye that closes from the wrong side. Common mode conversion shows up as an asymmetric eye, as increased electromagnetic emissions, or as a link that works on a short cable and fails on a longer one. Our article on PCB jitter sets out how those effects are measured and what the layout contribution looks like.

FAQ

• How close should a differential pair be routed? Close enough to meet the impedance target, and never so close that the spacing has to change repeatedly along the route. Consistency matters more than tightness.
• Does the gap have to stay constant? Yes, that is the rule that matters most. If the gap must change, the trace width should change with it to hold the impedance.
• Can the two traces be routed on different layers? Only for a short distance with a deliberate transition. Over any real length it destroys the coupling and the impedance control.
• Is length matching to the picosecond necessary? Only for the intra pair budget at the data rate in use. Inter pair matching on a bus is usually far more relaxed.
• Do differential pairs need ground vias at every transition? They need a defined return path. A ground via next to the signal vias is the standard way to provide one.

Summary

A differential pair works because the receiver subtracts the common mode component, and almost every routing rule exists to prevent the pair from generating common mode in the first place. Constant spacing preserves the impedance, matched length preserves the timing, and a continuous reference preserves the return path. Those three properties are worth more than any amount of cosmetic symmetry.

The practical method is to route the pair as a single object, to keep the geometry consistent through every transition, to correct any mismatch where it occurs, and to verify the result with an impedance measurement and a skew measurement rather than by inspection. Where a design is at the edge of its budget, the stackup and the via structure belong in the same conversation, and it is worth confirming with the fabricator what impedance tolerance can actually be held before the routing is committed to a number the process cannot deliver.