Introduction
A through hole via on a multilayer board spans the whole stackup, while the signal usually changes layer on only two or three of those layers. The unused barrel below the connection point is the stub. It is irrelevant at low frequency, but in a link running above ten gigabits per second it becomes an unterminated transmission line that resonates at a quarter wavelength and carves a deep notch into the channel response. This article starts from how the stub length is calculated, quantifies the relationship between resonant frequency and insertion loss, explains the controlled depth re-drilling process of back drilling and the constraints between bit diameter and remaining stub, analyses how depth tolerance and layer to layer registration turn a design target into a distribution across the panel, compares the three alternatives of blind and buried vias, stackup scheduling and simply not drilling, and closes with an actionable decision threshold and verification method.
1. Where the stub comes from: the unused barrel below the connection
A through hole via is manufactured by drilling through the whole board and then plating copper chemically and electrolytically, so the barrel is a metal tube spanning every layer. The signal changes layer at two specific layers and makes electrical contact with the barrel only there, while the remaining sections are plated but never connected into the signal path. The barrel below the connection point is the stub, its far end is open circuit, and electrically it is a branch line hanging off the signal path.
At low frequency the capacitive loading of that branch is negligible, but as the frequency rises to where the stub approaches a quarter wavelength, the open end is transformed into something close to a short at the junction and the energy at that frequency is absorbed and reflected. The channel response then shows a distinct notch and the eye closes around that frequency. Harmonics appear at odd multiples, the third, the fifth and so on, producing a train of notches.
One typical scenario shows the scale. On a three millimetre thick twenty four layer server board with the signal routed on layer three, a through hole transition leaves a stub of roughly two point eight millimetres. On a conventional laminate with a dielectric constant near four, the first resonance lands around thirteen gigahertz, squarely inside the Nyquist bandwidth of a high speed serial link. If a fifty six gigabaud four level pulse amplitude channel carries a loss budget of only twenty five to thirty decibels, that single via can consume three to six decibels of it.
2. Quarter wave resonance: converting stub length into resonant frequency
The fundamental resonant frequency of a stub has a usable approximation. It equals the speed of light divided by four times the stub length times the square root of the effective dielectric constant. The expression links two design variables directly, because a shorter stub raises the resonant point and a higher dielectric constant lowers it. In practice it is more useful the other way round. Fix the highest frequency of interest for the link first, then back out the longest stub that can be tolerated.
Run through the numbers on a conventional laminate and a four millimetre stub resonates first near nine point four gigahertz, three millimetres near twelve point five, two millimetres near eighteen point eight, one millimetre near thirty seven point five and half a millimetre near seventy five. A stub of zero point two millimetres resonates around one hundred and eighty seven gigahertz, outside the bandwidth of every protocol in use today. That is why the industry threshold is what it is. Evaluate once the stub exceeds fifteen mils and treat it as mandatory once it passes forty mils.
One point is often missed. The stub does not just carve a notch, it raises insertion loss across the whole band. A two millimetre stub adds roughly three to six decibels of loss at twelve point five gigahertz, a one millimetre stub about one to two, and after back drilling to zero point two millimetres less than zero point three. On a backplane channel that crosses four via transitions, leaving the stubs untreated can close the eye completely.
3. How back drilling works: controlled depth re-drilling, bit diameter and remaining stub
Back drilling is a second operation after plating, where a bit slightly larger than the original hole re-enters the board from the opposite face and grinds away the plated copper of the stub. The bit is typically zero point two to zero point three millimetres larger than the original through hole, which guarantees complete copper removal along the stub while leaving the barrel at the connection layer intact. That diameter difference is what forces a wider annular ring on the entry side pad and forces routing around the via to keep clear.
Why leave any stub at all? If the plating were stopped exactly at the target layer, the combination of axial depth error and layer to layer registration error would easily drill through the connection and destroy the joint that has to stay. The design therefore reserves a safety distance above the target layer, and production typically lands on a remaining stub of zero point fifteen to zero point twenty five millimetres. That residual pushes the resonance above one hundred gigahertz and removes it from any practical bandwidth, but it is also where the yield risk lives.
Three process parameters decide the result. The first is depth control, where mechanical stops, spindle current sensing and laser measurement are the options, and volume production generally holds plus or minus zero point one millimetre with premium capability reaching plus or minus zero point zero five. The second is bit condition and hit count, because bit wear raises exit burrs and surface roughness. The third is registration, since misalignment between the second drill pass and the original hole produces a skewed cut that leaves the remaining stub short on one side and long on the other, and it is the long side that dominates the resonance.
4. Tolerance stack up: how a design target becomes a distribution across the panel
A drawing that specifies a zero point fifteen millimetre stub will measure as a distribution from zero point zero five to zero point twenty five millimetres across a panel. That spread comes from three error terms stacking up. Back drill depth tolerance, layer to layer registration tolerance, and the tolerance of the original hole depth together with the laminate thickness. The first two typically contribute plus or minus zero point one to zero point one three five millimetres each, and the third depends on the laminate thickness tolerance and the actual dielectric thickness after lamination.
The consequence is worse than it looks. A zero point fifteen millimetre stub resonates near one hundred and fifty gigahertz while zero point twenty five millimetres drops the figure to roughly ninety gigahertz, a difference of a whole usable band. On a product running above fifty gigahertz that spread shows up as measurable variation in channel response between boards from the same batch, appearing in production test as margin that comes and goes. It is also why high end designs prefer to push the target stub shorter rather than use the full allowance.
One further detail is the symmetry of differential vias. If the two vias of a pair differ in path length by five to ten mils, both odd mode and even mode impedance shift, the differential to common mode and common mode to differential conversion terms rise, and common mode noise goes up with them. On critical high speed differential paths the symmetry tolerance between the two holes should stay within plus or minus five mils. Back drilling does nothing to fix asymmetry, so if the layout is not aligned to begin with the problem survives the back drill.
5. Three alternative paths: blind and buried vias, stackup scheduling, and simply not drilling
The first alternative is blind and buried vias. A blind via produces no stub by construction because it connects only the layers that are needed, and it also lowers via inductance, so both insertion loss and consistency beat a back drilled hole. The price is sequential lamination, where each extra lamination cycle adds cost, lead time and registration risk. Above twenty five gigahertz, or wherever the required remaining stub is extremely tight, the electrical consistency of a blind via is usually worth the premium.
The second is stackup scheduling, and it is the cheapest of the three. Placing high speed signal layers near the surface makes the stub short by itself, and placing critical layer transitions near the middle of the board can halve it. This path adds no process step at all and only requires that the layer transition positions of the high speed nets be planned during stackup design. Its limit is that total layer count and connector position constrain the outcome, so not every net can be moved to an ideal layer.
The third is simply not drilling, provided the arithmetic has been done. If the stub length is already inside the threshold, adding a back drilling step brings no benefit and only adds cost and yield risk. The way to decide is to compute the stub length of every high speed via from the stackup, estimate the resonant frequency, and compare it against the Nyquist frequency and the loss budget of the link. That calculation has to happen before the stackup is frozen rather than after the board is released. A hybrid is also common, using back drilled through vias for most routing and microvias only in the dense ball grid area.
6. Decision thresholds and how to verify
Reduced to rules, the quantitative relationships give three tiers. At twenty five gigabits non return to zero or fifty six gigabaud four level pulse amplitude and above, with a board thicker than two point five millimetres and a stub longer than one millimetre, back drilling is mandatory. Between ten and twenty five gigabits, or where the board exceeds one point six millimetres with signals on outer layers running through thick vias, back drilling is usually required. Below five gigabits on boards thinner than one point six millimetres the stub is inherently short and normally needs no treatment.
Verification runs in two steps. The first is a microsection, which measures the remaining stub directly and also checks the enlarged entry, whether the plating transition shows a crack or a separation, and the exit burr. A section is destructive and a product board cannot be sacrificed, so the second step uses a test coupon. Build a coupon at the board edge containing a comparable via, measure it with time domain reflectometry or a vector network analyser, and compare against the simulation, because the difference is the evidence for whether the design assumption about the stub holds.
Three documents have to be ready before release. The drill chart must state clearly which holes are back drilled, from which face, and the target remaining stub. The stackup drawing must show the registration relationship between the back drilled layers and the mechanical layers. The impedance control requirement matters because back drilling changes the reference structure around the via, and impedance is typically held to plus or minus five percent. Without these three, the fabricator can only guess from experience and the probability of a rework cycle rises sharply.
Figure Notes
Figure 1｜Stub length against first resonant frequencyFigure 2｜Additional insertion loss at twelve point five gigahertz by stub lengthFigure 3｜Back drilled through via versus blind and buried viaFigure 4｜The back drilling process flow and its control pointsFigure 5｜Five inputs that must be frozen before laying out for back drilling