Differential Pair: Preparation, Placement and Process Control

A differential pair carries a signal on two conductors and reads the difference between them. That arrangement rejects common mode noise, halves the effective switching current and gives a predictable impedance, but only if the two traces stay coupled and matched along the whole path. Routing practice is what preserves those properties.

What a Differential Pair Actually Does

The receiver responds to the voltage difference between the two lines, so noise that appears equally on both is subtracted out. The pair also radiates less because the two opposing currents produce fields that largely cancel. Both benefits depend on the pair remaining balanced, because imbalance converts differential signal into common mode noise.

The impedance that matters is the differential impedance, which is the impedance seen by the differential signal. It depends on the trace geometry, the dielectric and the spacing between the two traces, so a change in spacing changes the impedance even when each trace individually keeps its width.

Impedance Control for Pairs

The stack-up should be designed so that the target differential impedance can be achieved with a practical width and gap. Width controls the single ended impedance and the gap controls the coupling, so the two parameters are adjusted together until both the differential impedance and the physical fit are acceptable.

Tolerance is inherited from the same sources as single ended impedance: dielectric thickness, copper thickness and etch variation. Because the gap is usually small, its tolerance matters proportionally more, which is why pairs are often routed with a slightly wider gap than the minimum the process allows. The stack-up considerations are described in this guide to high speed design rules.

Keeping the Pair Coupled

The pair should be routed together with a constant gap for its whole length. When the two traces separate, the coupling changes, the impedance rises locally and a reflection is created. Where a pair must pass around an obstacle, the deviation should be made with both traces together rather than by detouring one of them.

Short separations are tolerable at moderate data rates, but the acceptable length depends on the rise time of the signal. A useful rule is that the disruption should be short compared with the distance the signal travels during one rise time, which is a small fraction of a millimetre at multi gigabit rates.

Skew and Length Matching

Skew is the difference in arrival time between the two halves of the pair. It converts part of the differential signal into common mode noise and reduces the eye opening at the receiver. Skew comes from differences in length, from different via counts, from different layer transitions and from the glass weave of the laminate.

Length matching compensates for the geometric differences by adding small serpentine adjustments to the shorter trace. The compensation should be placed close to where the mismatch occurs rather than at the end of the route, because the skew is present for the whole length between the mismatch and the correction. One cause of skew that is often overlooked is the glass weave of the laminate, which makes the propagation velocity slightly different for traces running at different angles to the weave, as described in this guide to laminate material properties.

Vias, Layer Changes and Returns

Every via is a discontinuity. A pair that changes layer should do so with vias placed symmetrically so that both traces see the same structure, and the return path through the reference planes has to be continuous. A pair that crosses a plane split loses its return path and radiates.

Ground stitching vias placed near the signal vias give the return current a short path between planes. Without them the return current must find its own way around, which increases loop area and degrades the very noise performance the pair was chosen to provide.

Termination and Common Mode

Termination should match the differential impedance and, where the receiver requires it, provide a path for common mode current. A pair terminated only differentially leaves common mode energy to reflect, which shows up as radiated emissions even when the differential signal is clean.

The common mode termination is usually a capacitor to ground at the receiver, or a resistor network specified by the interface standard. Following the standard is safer than improvising, because the value interacts with the driver and with the connector.

Routing Near Other Signals

A pair couples to whatever is near it, so spacing from other nets matters. The most important rule is to keep other traces away from the pair by at least the same distance the pair uses internally, and to avoid routing a single ended signal between the two halves of a pair, which disrupts the coupling completely.

Where a pair runs alongside another pair for a long distance, the two can couple to each other, which transfers noise from one link to another. Increasing the spacing between pairs, or routing them on different layers, reduces the effect. Crosstalk behaviour in general is described in this guide to high speed routing practice.

Test Structures and Verification

Differential impedance is verified on a coupon that includes a pair with the production geometry, measured with a time domain reflectometer or a network analyser. Measuring the impedance of a single trace and assuming the pair is correct is a common error, because the differential mode sees a different structure.

The coupon should be built into the panel border so that it experiences the same lamination and plating. Where the design uses several pair geometries, each should be represented, because a stack-up that works for one width and gap may not work for another that is used elsewhere on the board. A test coupon in the panel border is the practical way to make that measurement part of every lot rather than a one off exercise.

Common Layout Mistakes

The most frequent mistake is splitting the pair to route around a via or a component, which destroys the coupling over that length and creates an impedance discontinuity. The second is treating the pair as two independent traces and length matching them separately, which produces a matched but uncoupled route with none of the intended benefits.

A third is forgetting the return path. A pair that crosses a plane split, or that changes reference planes without stitching vias, has a broken return and will radiate. A fourth is adding excessive serpentine compensation for skew, which adds length and discontinuities in exchange for a match that the interface did not require.

FAQ

Do differential pairs need to be length matched exactly? They need to be matched to within the skew budget set by the interface and the rise time, which is often a small fraction of the bit period rather than exactly zero. Matching far beyond the requirement wastes routing area and can introduce its own discontinuities.

Can the two traces of a pair be routed on different layers? They can, and some designs do it, but the two traces then see different dielectrics and different reference planes, which introduces skew and impedance mismatch. Where it is unavoidable, the length and via structure should be matched as closely as possible.

Does the pair need a ground plane underneath? A solid reference plane is what makes the impedance predictable and provides the return path. A pair routed over a split or over a plane with large slots will not hold its impedance and will radiate, whatever the trace geometry says.