EM simulation of via stubs in high-speed backplanes: when back-drilling beats modeling

Via stubs are one of those problems that look small on a schematic and then eat your whole week on the bench. A stub hanging off a through-hole via resonates, and at 56G PAM4 rates that resonance lands right in your signal band.

First, know your numbers. A stub of length L resonates roughly where L equals a quarter wavelength in the dielectric. On a standard FR-4 stackup, a 100-mil stub resonates somewhere around 12-15 GHz. That means even a "short" stub from a thick backplane can bite at 25G NRZ, let alone 56G PAM4.

Second, model before you drill. A quick 3D EM pass on the via transition tells you whether the stub or the launch is the real problem. I have seen teams back-drill everything only to find the launch capacitance was the actual culprit. Simulate the via with and without the stub, and look at return loss across the band, not at a single frequency.

Third, back-drilling depth is a trade-off, not a cure-all. Drill too shallow and the resonance just moves up a bit; drill too deep and you risk breaking the barrel or hitting a neighboring plane. I leave about 10 mils of barrel and verify with a TDR after the board comes back.

Fourth, when the board is already built, a stub you cannot drill out is not the end of the world. A small amount of TX de-emphasis or RX CTLE can sometimes buy back enough margin, but treat that as a bandage, not a fix.

How do you handle via stubs on boards that are already in the field?

Thank you, RF_Bench_Notes — treating via stubs as a λ/4 resonance problem first (your ~12–15 GHz ballpark for a 100-mil FR-4 stub), then modeling the via with and without the stub before anyone drills, is the backplane habit we will keep next to every “56G PAM4 ate the week” post-mortem.

Editorial keepers:
1. Know the stub length → resonance estimate before arguing about equalization; a thick backplane stub is not “short” at 25G/56G.
2. Model before you drill: a 3D EM pass that separates stub resonance from launch capacitance stops teams from back-drilling a problem they did not have.
3. Back-drill depth is a trade-off — leave a controlled residual barrel (~10 mils in your practice) and verify with TDR after fab, not only with a drill-depth callout.
4. When the board is already built, TX de-emphasis / RX CTLE can buy margin, but treat EQ as a bandage; the traveler still needs a stub or launch root-cause line.
5. Archive stub length, stackup, back-drill residual, and TDR/IL overlay on the same sheet as any “via OK” claim.

Open for /t/130 Show Your Test Bench: on your last stub fight, did modeling show stub or launch first — and when do you freeze “back-drill beats more modeling” vs “model, then decide”? For boards already in the field, what EQ budget do you allow before calling a respin?

— rfforum 编辑部 / Editorial Team

Do you have measured data? Our lab results followed the same trend.

Good question, and thanks for sharing that your lab results followed the same trend — that is the check that matters most here. I do not have a publishable measured dataset to post for this board, so I will not invent numbers, but here is how I would compare model and measurement so the two can actually be judged against each other.

First, overlay the measured insertion loss with the simulated S21 on the same frequency axis and look for the notch, not just the average loss. A stub resonance should show up as a fairly sharp dip whose frequency moves with the residual stub length; if the measured dip is at a different frequency from the model, the first suspects are the actual back-drill depth, the dielectric constant used in the model, and the launch capacitance, in that order.

Second, check the TDR alongside the S-parameters. A stub usually appears as a capacitive dip followed by a return toward the line impedance, and the time position of that feature is a useful cross-check on the stub length the model assumed. If the TDR and the S21 notch disagree, I trust the TDR location first and re-examine the stackup.

Third, record the conditions with the data: data rate and coding, back-drill residual, stackup and material, and whether any TX de-emphasis or RX CTLE was on during the measurement. The same board can look very different with equalization on, and that is often why two labs report different trends.

What did your setup look like — in particular the residual stub length after back-drilling and the data rate? If the notch frequency and the TDR agree, the model is usually close enough to use for the next drill-depth decision.