Note 17Signal integrity Rev 1 Watch the video · 0:09

Via stubs

How long a via stub can a fast link take?

About 23.6 mm divided by the lane’s rate in GBd: 2.3 mm for 10GBASE-KR, 0.7 mm for PCIe 5. Past that, the stub costs more than about 0.5 dB at the link’s Nyquist frequency. A via to layer 3 of a 1.6 mm board leaves 0.94 mm, fine to about 25 Gb/s.

The figure sends one fast pulse into two vias that drop from the top layer to layer 3. In the first, the plated barrel runs on to the bottom of the board. That leftover length, the stub, rings and sends 13% of the pulse back.

In the second, a drill has cut the stub away from below. Only 0.9% comes back.

A pulse meets the via

20 ps wide, slowed 50 billion times

Pulse going in
Echo coming back

Above each via, the voltage along 9 mm of top trace into it and 9 mm of layer 3 out of it. Below, the via itself in section through the MNT Reform motherboard 3.0, to scale, lighting up where the wave passes. Sent back counts the share of the pulse’s energy that returns up the top trace.

The numbers

Via	Energy sent back	Energy through	Stub	Down the stub and back	Peak on layer 3
Through all six layers	13%	87%	0.94 mm	13.4 ps	0.82 V
Back-drilled	0.9%	99%	0.2 mm	2.8 ps	0.99 V

What a via stub is

One via on the MNT Reform motherboard 3.0

A via carries a signal from one layer to another through a plated hole. Drilled straight through the board, its barrel runs from the top layer to the bottom, whichever layers the signal uses.

When the signal leaves on an inner layer, the barrel below it carries nothing on. That leftover length is the stub. It hangs off the line like a short wire with an open end. A TDR sees it as a dip, as Note 16 shows.

On a 1.6 mm, 6-layer board, a via from the top layer to layer 3 leaves 0.94 mm of stub.

The via, cut in half

A 0.4 mm drill with 0.8 mm pads on every layer, to scale. The planes are cleared 0.3 mm around the pads. The line shows the signal’s way through.

The numbers

Layer	What it is	Depth of its center
F.Cu	Top	0.018 mm
In1	Ground	0.327 mm
In2	Layer 3	0.635 mm
In3	Power	0.945 mm
In4	Ground	1.254 mm
B.Cu	Bottom	1.562 mm
Stub	Barrel below layer 3	0.94 mm

The MNT Reform laptop’s motherboard has vias like this on its PCIe lines. The processor module reaches the M.2 slot over two pairs, one each way.

Each pair leaves the module’s socket on short top-layer traces, drops to layer 3 and comes back up at the slot. All eight vias are plated through all six layers, so each one carries a 0.94 mm stub.

Where the pairs run

Top layer
Layer 3

The whole board from the top, and the box around the pairs enlarged. The copper is the top layer’s, read from the board’s design files. The arrow points at the via in the sections above.

The numbers

Pair, line	End	Drill	Pad	Plated
PCIe RX N	M.2 slot	0.4 mm	0.8 mm	F.Cu to B.Cu
PCIe RX P	M.2 slot	0.4 mm	0.8 mm	F.Cu to B.Cu
PCIe TX N	M.2 slot	0.4 mm	0.8 mm	F.Cu to B.Cu
PCIe TX P	M.2 slot	0.4 mm	0.8 mm	F.Cu to B.Cu
PCIe RX N	Processor module	0.4 mm	0.8 mm	F.Cu to B.Cu
PCIe RX P	Processor module	0.4 mm	0.8 mm	F.Cu to B.Cu
PCIe TX N	Processor module	0.4 mm	0.8 mm	F.Cu to B.Cu
PCIe TX P	Processor module	0.4 mm	0.8 mm	F.Cu to B.Cu

A quarter-wave notch

Three stubs, swept to 60 GHz

An edge that runs down the stub bounces off its open end and comes back up. At the frequency where the stub is a quarter wave long, the bounce arrives back exactly out of step. The open end then looks like a short at the junction, and almost nothing gets past.

f_notch=c4Lε_r

=300 mm/ns4 × 0.945 mm × 2.12=37.4 GHz

So the notch falls as one over the stub’s length. In FR-4 it sits near 35 GHz divided by the stub in mm.

How much gets through

This board, 0.94 mm
3.2 mm board, 2.6 mm
Back-drilled, 0.2 mm

S21 of the via alone, from 12 mm of top trace to 12 mm of layer 3. The model has no loss, so its notches have no floor. They are cut off at −30 dB.

The numbers

S21 at	Back-drilled	This board	Thick board
Stub	0.2 mm	0.94 mm	2.6 mm
5 GHz	−0.01 dB	−0.08 dB	−0.56 dB
10 GHz	−0.03 dB	−0.32 dB	−4.00 dB
12.5 GHz	−0.04 dB	−0.51 dB	−11.85 dB
16 GHz	−0.06 dB	−0.89 dB	−7.22 dB
20 GHz	−0.09 dB	−1.54 dB	−1.27 dB
28 GHz	−0.16 dB	−4.42 dB	−0.04 dB
37.4 GHz	−0.24 dB	−56.73 dB	−3.55 dB
50 GHz	−0.33 dB	−2.32 dB	−0.43 dB
Notch	176.7 GHz	37.4 GHz	13.8 GHz

338

Quarter-wave notch

This board’s stub notches at 37.4 GHz

Field marks. S21 falls away to a deep notch at 37.4 GHz and comes back above it. Back-drilled to 0.2 mm, the same via loses under 0.3 dB all the way to 40 GHz, and its notch moves out to 177 GHz.

Look-alike. A 3.2 mm board leaves 2.6 mm of stub under the same via. It notches at 13.8 GHz, and again at three times that, 41 GHz.

Range. Any via whose barrel runs on past the layer the signal leaves on. Half the stub, twice the frequency.

339

Long before the notch

The longest stub for six links

A link never runs at the notch. What matters is its Nyquist frequency, half its baud rate: the tone of its fastest pattern, 1010. That is where a channel’s loss is usually judged.

The stub already costs about 0.5 dB at a third of the notch and 1 dB at 45% of it. That held for every stub from 0.5 to 4 mm. Keep the notch three times above Nyquist, and the stub must be shorter than

L_max=c6ε_rR

=300 mm/ns6 × 2.12 × 10.3 GBd=2.3 mmfor 10GBASE-KR

That is 23.6 mm divided by the rate in GBd.

The longest stub each link takes

Before the stub costs about 0.5 dB at the link’s Nyquist frequency. GBd is per lane. PCIe 6 and 112G send four levels a symbol, PAM4. The rest send two, NRZ.

The numbers

Link	Rate a lane	Longest stub	This board fits	Thick board fits
10GBASE-KR	10.3 GBd	2.28 mm	Yes	No
PCIe 4	16 GBd	1.47 mm	Yes	No
USB4, DP UHBR20	20 GBd	1.18 mm	Yes	No
25G Ethernet	25.8 GBd	0.91 mm	No	No
PCIe 5, PCIe 6	32 GBd	0.74 mm	No	No
112G PAM4	53.1 GBd	0.44 mm	No	No

340

Stub limits

This board’s stub stops just short of 25G Ethernet

Field marks. The 0.94 mm stub fits every link up to USB4’s 1.2 mm. 25G Ethernet takes 0.91 mm, so it misses by 0.03 mm.

Look-alike. The same via on a 3.2 mm board leaves 2.6 mm. That is past even 10GBASE-KR’s 2.3 mm.

Range. PCIe 6 runs at PCIe 5’s 32 GBd, so it takes the same 0.7 mm. 112G PAM4 at 53 GBd takes 0.4 mm, which only back-drilling or a blind via gets under.

341

Thicken the board yourself

1.6 to 3.2 mm, plated through or back-drilled

Backplanes and big server boards often run 3 mm thick or more. The via to layer 3 stays the same, but the barrel below it grows with the board.

Slide the board thicker and watch the fastest lane the stub allows fall. Then back-drill it.

Board

1.6 mm

Via

Via stubs

1.6 mm board, plated through: under 0.5 dB to 24.9 Gb/s

Within reach
Too fast

Stub: 0.94 mm
Notch: 37.4 GHz
Fastest lane: 24.9 Gb/s

The bar is the fastest NRZ lane whose Nyquist frequency the stub costs under about 0.5 dB. As the board thickens, the layers below layer 3 spread out evenly. Back-drilled, 0.2 mm is left on every board, good to 118 Gb/s, past the end of the scale.

The numbers

Board	Via	Stub	Notch	Fastest lane
1.6 mm	Plated through	0.94 mm	37.4 GHz	24.9 Gb/s
1.6 mm	Back-drilled	0.2 mm	176.7 GHz	117.8 Gb/s
1.8 mm	Plated through	1.2 mm	30.3 GHz	20.2 Gb/s
1.8 mm	Back-drilled	0.2 mm	176.7 GHz	117.8 Gb/s
2 mm	Plated through	1.4 mm	25.9 GHz	17.3 Gb/s
2 mm	Back-drilled	0.2 mm	176.7 GHz	117.8 Gb/s
2.2 mm	Plated through	1.6 mm	22.6 GHz	15.1 Gb/s
2.2 mm	Back-drilled	0.2 mm	176.7 GHz	117.8 Gb/s
2.4 mm	Plated through	1.8 mm	20.0 GHz	13.3 Gb/s
2.4 mm	Back-drilled	0.2 mm	176.7 GHz	117.8 Gb/s
2.6 mm	Plated through	2.0 mm	18.0 GHz	12.0 Gb/s
2.6 mm	Back-drilled	0.2 mm	176.7 GHz	117.8 Gb/s
2.8 mm	Plated through	2.2 mm	16.3 GHz	10.9 Gb/s
2.8 mm	Back-drilled	0.2 mm	176.7 GHz	117.8 Gb/s
3 mm	Plated through	2.4 mm	14.9 GHz	10.0 Gb/s
3 mm	Back-drilled	0.2 mm	176.7 GHz	117.8 Gb/s
3.2 mm	Plated through	2.6 mm	13.8 GHz	9.2 Gb/s
3.2 mm	Back-drilled	0.2 mm	176.7 GHz	117.8 Gb/s

The eye on a thick board

25 Gb/s, three vias compared

The decibels show up as closed eyes. These are 25 Gb/s of random bits after 5 cm of FR-4 and the via, with no equalization. Note 8 shows how an eye is built.

The eye after the via

Every bit laid over the next, two bits wide. The darker the shade, the more bits passed there. The bar is the opening at the best moment to sample, as a share of the full swing.

The numbers

Eye open at	1.6 mm board	3.2 mm board	3.2 mm, back-drilled
Stub	0.94 mm stub	2.6 mm stub	0.2 mm left
10 Gb/s	79%	78%	79%
16 Gb/s	71%	61%	72%
25 Gb/s	58%	4.7%	62%

342

Eye after the via

A 2.6 mm stub closes the eye to 4.7%

Field marks. On the 3.2 mm board the stub’s echo comes back 36 ps later, about one bit at 25 Gb/s. It lands on the next bit, each rail splits in two, and the eye is 4.7% open.

Look-alike. The 1.6 mm board’s 0.94 mm stub leaves it 58% open. Back-drilled, the 3.2 mm board opens to 62%, a little better than the thin board, since only 0.2 mm is left.

Range. At 10 Gb/s all three are 78 to 79% open. At 16 Gb/s the thick board falls to 61% and the others stay above 70%.

343

What each fix buys

The same via, four ways

There are four ways to deal with a stub, cheapest first. Leaving it costs nothing. Routing the pair to the far layer costs nothing either, if the layer is free. Back-drilling adds a drilling step from the far side. A blind via needs a board built in stages, which is worth it only if the board needs that anyway.

Four ways through

Each section is the via drawn to scale, with the signal’s way through it. Routed to the far layer, the whole barrel carries the signal.

The numbers

	Fix	Stub	Notch
1	Leave it	0.94 mm stub	37.4 GHz
2	Far layer	No stub	None
3	Back-drill	0.2 mm left	177 GHz
4	Blind via	No stub	None

344

Key to via stubs

Does this via need a fix?

1aStub shorter than 23.6 mm ÷ the lane’s GBdLeave it
1bStub longer than thatGo to 2
2aThe pair can leave on the far layerRoute it there
2bIt has to leave on an inner layerGo to 3
3aThe board is built in stages anywayA blind via
3bA standard boardBack-drill, 0.2 mm left

This board’s 0.94 mm stub passes 1a up to 25 GBd.

345

Conditions

Board: MNT Reform motherboard 3.0, 6 layers, 1.58 mm. Its design files are open hardware under CERN-OHL-S v2.
Stack-up: Six 35 µm copper layers and five 0.274 mm FR-4 layers, εr 4.5. The top trace sees 3.4 with the air above it.
Via: 0.4 mm drill, 0.8 mm pads on every layer, planes cleared 0.3 mm around the pads. Top layer to layer 3: 0.62 mm of barrel used, 0.94 mm left below.
Where: The PCIe pairs from the processor module’s socket, U1, to the M.2 slot, J10
Model: Every piece is a transmission line: 12 mm of 50 Ω top trace in, the used barrel as a 45 Ω line, the stub as an open 45 Ω line, 12 mm of 50 Ω layer-3 trace out. Each pad on a layer the signal doesn’t use adds 0.05 pF to ground. Both ends matched. No loss in the via.
Solved: In time, by finite differences, for the pulse. In frequency, with chained ABCD matrices, for S21, the limits and the eyes. The two agree within 0.1 dB up to 34 GHz.
Pulse: 1 V, 20 ps wide at half height
Eyes: NRZ, 1500 random bits, edges 0.35 of a bit long, 5 cm of FR-4 with a loss tangent of 0.02 before the via, no equalization
Links: Each standard’s published rate per lane, in GBd
Thicker boards: The layers below layer 3 spread out evenly. Everything above stays as built.

Simulated only. Nothing on this page has been measured on a board yet.

Limits

A line model, not a 3D field solve. The barrel’s 45 Ω and the pads’ 0.05 pF are set by hand, not solved, so the notch could sit a few GHz away.
One line on its own. The pair’s other line, the ground vias beside it and the space between the planes are left out. The RX vias have no ground via within 1.6 mm.
With no loss, the notches go all the way down. Real ones are shallower and a little wider.
The eyes are this via’s alone. A real link adds its connector, its package and its equalizer.

Files

viasim/line1d.pyThe pulse in time
viasim/freq.pyS21, the notch, the limits and the eyes
layout/site_board.pyWhere the pairs run, read from the board file

Sources

MNT Reform

Like this note?

Rev 1 · 5 Nov 2026

Rev 15 Nov 2026First published

Each fix gets a line here, with the reader's name if they want credit.

Did it answer your question?

Want more on this?

Report a problem

What should the next note be? Ask for one