Via Optimization For High Speed Channels

A via is a short vertical section of transmission line with a length that depends on the board thickness and a geometry that no two dimensional model describes well. On a slow net it is a connection; on a fast channel it is a discontinuity that contributes loss, reflection and, in the worst case, a resonance that removes a band of frequencies entirely.

This article covers via optimization in the order the parameters should be adjusted, and the point at which the remaining discontinuity is small enough to ignore.

The starting point is to know what the via is doing. Measuring or simulating the transition before optimising it prevents effort being spent on a parameter that does not control the result.

What Makes A Via Discontinuous

Two effects dominate. The barrel presents capacitance to every plane it passes, and the unused portion of the barrel below the layer where the signal leaves forms an open stub that behaves as a resonator. The capacitance lowers the local impedance; the stub creates a notch at the frequency whose quarter wavelength matches its length.

On a thin board with few layers, both effects are small and the via is nearly transparent. On a thick backplane the stub can be several millimetres long, and the notch can fall inside the band of a multi-gigabit link, which is why backplanes are the application where via treatment matters most. The general via design rules cover the manufacturability of the structure; the electrical treatment is a separate exercise.

Antipad And Pad Size

The antipad on each layer controls the capacitance between the barrel and the plane. Enlarging it reduces the capacitance and raises the local impedance, which helps where the via is capacitive. Enlarging it also weakens the plane and lengthens the return current path, which adds inductance, so there is an optimum rather than a direction.

The capture pad on the layer where the trace connects also contributes capacitance, and its size is usually set by the fabrication rules. Where the pad can be reduced within those rules, doing so reduces the capacitance without the penalty of a larger antipad, which is why pad size is the first parameter to examine.

It is worth separating the two failure mechanisms because their remedies differ. A capacitive via that lowers the local impedance is best treated by reducing the pad and enlarging the antipad, which reduces the capacitance. A resonant stub that notches the band can only be treated by removing the stub, and no adjustment of the antipad will shift the notch frequency far enough to matter.

Removing The Stub

Back drilling removes the unused portion of the barrel after plating, which eliminates the stub and removes the notch. It is the single most effective treatment on a thick board, and its cost is a process step and a depth tolerance that has to be controlled so that the drill does not remove the connection it is meant to preserve.

Where back drilling is not available, the alternative is to plan the layer assignment so that the signal travels from the outer layer to a plane near the far side, minimising the unused length. Blind and buried vias achieve the same result by construction, and the selection of the via stack is therefore an electrical decision as well as a density one.

Ground Return Vias

A via transition needs a return path, and the return path is a ground via placed close to the signal via on the same layer transition. Without it, the return current has to find its own route through the plane, which lengthens the loop and adds inductance to the transition.

The placement is more important than the number. A ground via within a fraction of a millimetre of the signal via couples most of the return, while several ground vias at a distance of a millimetre help much less. On a differential pair, the ground vias should be placed symmetrically so that both members of the pair see the same environment.

When The Via Is Small Enough To Ignore

The criterion is the same as for any discontinuity: the via matters when its electrical length is a significant fraction of the rise time distance. A via a fraction of a millimetre long on a thin board is negligible for edges measured in hundreds of picoseconds, and the same via on a four millimetre thick backplane is not.

In practice the check is to compare the delay through the via with the rise time of the signal. Where the ratio is small, the via can be treated as a lumped capacitance and the treatment is limited to choosing a sensible antipad. Where the ratio is large, the treatment has to address the stub.

The layer stack also offers a lever that costs nothing. Placing the signal on a layer close to the layer it needs to reach, and choosing a stackup that keeps the plane pairs close together, reduces the length of every via on the board at once. This is one of the reasons a high speed design with a well chosen stackup needs less drastic via treatment than one built on a generic stackup.

A Sequence For Optimisation

Start with the layer assignment, because moving a signal to a layer closer to its destination plane may remove the need for any other treatment. Then reduce the pad size within the fabrication rules, then adjust the antipad, then add the return vias, and only then consider back drilling or a more complex via structure.

The sequence matters because the early steps are free and the later ones are not. Where the structure uses a filled via for assembly reasons, the filling process interacts with the electrical treatment, and the plating and filling of the barrel should be planned at the same time as the stub removal.

Verifying The Result

Simulation is the practical tool, using a three dimensional model of the transition with the actual stackup. The output that matters is the insertion loss of the transition and the return loss, and the two together show whether the via is acceptable. A stub resonance appears as a sharp feature in the insertion loss curve and is easy to identify.

Measurement is harder, because the via is inside the board and cannot be probed directly. A coupon that contains the same transition can be measured, and the comparison between a coupon with the via and one without isolates its contribution. That comparison is a reasonable final check before a design is committed to a thick board.

A final consideration is consistency. A channel with one optimised via and five untreated ones performs much like a channel with six untreated ones, because the worst transition dominates the result. The treatment should be applied uniformly to every transition on the channel, and the review should confirm that it was.

FAQ

Does every via need a ground via beside it? Every via on a fast signal should have a return path, and a nearby ground via is the cheapest way to provide one. On slow nets the return finds its own way through the plane at a small cost.

Is back drilling always worth it? On a thick board with fast signals it usually is, because the alternative is a notch in the band. On a thin board the stub is short enough that the cost is hard to justify.

Can a via be made to have no effect at all? No, but it can be made small enough that its effect is within the margin the design has. The objective is a transition that consumes an acceptable share of the budget, not one that is invisible.