Trace Skew in Differential Pairs: Causes and Fixes

Why a Pair Has to Arrive Together

A differential pair works because the two traces carry equal and opposite signals, and the receiver responds to the difference between them. That mechanism depends on the two signals arriving at the same time. Any difference in arrival time is skew, and skew converts part of the differential signal into a common mode signal that the receiver cannot reject.

The consequences are not subtle. Common mode content radiates, which shows up as an EMC problem. It also eats into the timing budget, narrows the eye, and degrades the effective signal to noise ratio at the receiver. On a link running at multi-gigabit rates, a small amount of skew that was irrelevant at lower speeds becomes a measurable contributor to a closed eye.

Skew is therefore one of the few layout parameters that has to be managed explicitly rather than monitored after the fact, because it cannot be corrected on the assembled board.

The Two Kinds of Skew

It helps to separate the two, because they have different causes and different fixes.

• Intra pair skew: a mismatch between the two traces of the same pair, measured in picoseconds. This converts differential signal into common mode, and it is the direct subject of length matching in layout.
• Inter pair skew: a mismatch between separate pairs that are supposed to be aligned, such as the lanes of a parallel bus or the clock and data relationship. The consequence is a timing margin reduction rather than common mode conversion.

Most of the daily layout work concerns intra pair skew, but ignoring inter pair skew on a parallel interface leads to a system that fails at speed despite every individual pair being well matched.

Where Skew Comes From

• Unequal trace length: the obvious cause. A pair that has to turn a corner, or that splits around an obstacle, accumulates a length difference if the geometry is not compensated.
• Via asymmetry: when the two traces of a pair transition through vias at different distances, or through vias of different depth, the electrical path lengths differ. A pair that changes layer with one trace taking an additional length of routing accumulates skew that no amount of length matching on the surface layers can fully remove.
• Reference plane discontinuities: if the two traces do not traverse the same reference structures, the effective propagation velocity differs even when the physical length matches.
• Dielectric inhomogeneity: on some laminates, the woven glass reinforcement creates regions of slightly different dielectric constant. Two traces running over different glass bundles see different velocities, which produces skew that varies with position on the panel.
• Bends and geometry changes: a trace routed at 45 degrees behaves slightly differently from one routed on a right angle, and serpentine compensation patterns have their own effect on propagation.

The via and dielectric effects are the ones most often missed. Length matching at layout addresses the first of the list, which is also the least difficult, and treats the others as negligible when they are not.

How Skew Becomes a Visible Problem

Skew does not simply delay one edge. Inside a pair, it converts part of the energy into a common mode component, which has two consequences.

• Radiation and susceptibility: a pair carrying a common mode component behaves like an antenna, both emitting and receiving. This is why a link that passes its functional test can still fail an EMC test.
• Loss of receiver margin: the differential signal amplitude at the sampling instant is reduced, which appears as a closure of the eye rather than as a clean error.

Measuring the effect directly is done with an eye diagram, and the contribution of the board can be separated from the transmitter and receiver behaviour by measuring the channel with a controlled stimulus. Skew appears in the eye as a reduction in the horizontal opening and as common mode energy in the spectrum.

Fixing Skew in Layout

• Match the length within the pair, and state the tolerance as a time rather than as a distance where possible. A skew budget of a few picoseconds corresponds to a very small length difference at typical propagation velocities, which is why matching to a generic rule of thumb is often not tight enough.
• Match at the point where the mismatch occurs. Compensation added far from the source of the mismatch does not cancel the timing error. If one trace of a pair runs around an obstacle, the compensation belongs near that detour.
• Keep the pair symmetric through bends. Both traces should turn together, with equal geometry, rather than one taking a longer inside path. This is where the standard practice of shifting the pair as a unit originates.
• Compensate with an appropriate pattern. Tight serpentine compensation has its own electrical character: the successive segments couple to each other and the propagation delay per unit length is not the same as a straight trace. Loose, spread out compensation is electrically better than a dense shaving pattern, even where it is less compact.
• Keep the gap constant. Changing the spacing to fit around a component changes the differential impedance as well as the coupling, so the gap is a controlled parameter rather than a layout variable.
• Preserve the reference. Both traces of a pair should see the same reference plane, with no split or plane boundary crossing under either one.

These are all layout decisions, and they have to be designed rather than checked at the end, because retrofitting a symmetric pattern into an already routed area is a redesign.

Vias and Layer Transitions

Via transitions are where a well matched surface routing can still develop skew.

• Keep the symmetry of the transition. Both traces should change layer through vias of the same depth and the same geometry, at the same point along the pair.
• Avoid adding a layer transition to only one trace. If one trace needs a via and the other does not, the two are no longer electrically equivalent, and the mismatch is not recoverable by surface length matching.
• Consider the stub effect. A via stub adds a small electrical discontinuity and a small delay, and because it is frequency dependent it does not affect the two traces identically unless the vias are identical. On high speed links this is one of the reasons back drilling and blind vias are used.
• Fan out symmetrically. The escape pattern from a connector or an area array package is a common source of one trace being longer than its partner; symmetric fanout geometry keeps them matched.

Manufacturing Contributions

Not all skew is designed in. Some arises from the fabrication process.

• Glass weave effects: on standard woven laminate, the glass bundles have a different dielectric constant from the resin rich regions between them. Two traces running over different weave patterns propagate at slightly different velocities. Randomising the phase of the weave relative to the routing, using a spread glass laminate, or accepting a slightly larger skew budget are the mitigations. On very high speed links this effect is a design consideration in its own right, not a detail.
• Etch uniformity: variation in trace width across the panel changes the propagation velocity slightly. Tight etch control, as used on any controlled impedance board, limits it.
• Dielectric thickness control: the same parameter that sets impedance also sets velocity, which is why fabrication control and signal integrity are the same conversation on high speed boards.

The practical implication is that a skew budget derived purely from layout geometry is optimistic. Allowance has to be made for the manufacturing contribution, and on tight designs the laminate choice is part of the skew strategy rather than only the loss strategy.

Verification

• Channel measurement: the time domain response of the pair, measured with a controlled stimulus, gives the delay difference directly.
• Eye diagram measurement: shows the combined effect of skew along with the other impairments in the channel.
• Impedance and loss verification on coupons, using TDR impedance testing, confirms that the manufactured geometry matches the design, which is a prerequisite for the skew prediction being valid.
• Simulation with the actual stackup, including the via models, rather than with an idealised pair. On short pairs, the vias dominate the skew calculation.

When Skew Matters Enough to Manage

Not every differential pair requires picosecond matching. The question is whether the skew budget is being consumed by the board or by the transmitter and receiver.

• Low speed differential signalling: matching within a generous tolerance is usually adequate, and tight compensation adds effort without benefit.
• Multi-gigabit serial links: the budget is tight and the receiver’s equalisation can only recover so much, so intra pair skew has to be controlled deliberately.
• Parallel buses and memory interfaces: inter pair skew between lanes matters as much as intra pair matching, and the budget is shared across the whole bus.
• Very short channels: counterintuitively, via and connector geometry can dominate the skew on a short link, so a short pair with asymmetric vias may be worse than a longer, symmetric one.

Frequently Asked Questions

What is intra pair skew? The difference in propagation delay between the two traces of a differential pair, usually expressed in picoseconds. It converts differential signal into common mode noise.

Why is skew a problem if the link still works? Because it reduces receiver margin and radiates. A link with skew may pass functional test and fail an EMC test, or perform reliably on the bench and fail in the field.

Does length matching solve skew? It addresses the geometry contribution, which is the largest and the easiest to control. Vias, reference plane discontinuities and dielectric inhomogeneity also contribute and are not corrected by matching the surface length.

Why is serpentine compensation not always the answer? Because densely packed compensation patterns couple to each other and propagate differently from a straight trace. The compensation is still appropriate, but the pattern should be loose and placed near the source of the mismatch.

How is skew verified on the finished board? By measuring the channel with a controlled stimulus and by eye diagram evaluation, supported by impedance and loss verification on coupons.

Summary

Trace skew is a timing problem with an electromagnetic consequence. Differential signalling depends on the two traces arriving together, and any difference in arrival time turns part of the signal into common mode energy that radiates and reduces the receiver’s margin.

The controllable contributors are unequal trace length, asymmetric vias, reference plane discontinuities and dielectric inhomogeneity in the laminate. Length matching addresses the first, and it is where most of the effort goes, but the other three mean that a well matched pair can still be skewed in practice. The fixes are structural: symmetric layer transitions, equal reference conditions, loose and correctly placed compensation, and a laminate choice suited to the speed of the link.

Because the cause can be a design decision, a fabrication variation or a laminate property, managing skew requires the design and the manufacturing process to be considered together. On links where it matters, the skew budget should be allocated explicitly between the board and the silicon, and verified by measurement rather than assumed from the layout.