RFrftools.io
Signal IntegrityMarch 1, 20268 min read

Eye Diagram Analysis for 10 Gbps SerDes: Validating Your Channel Before You Spin

A PCB designer routes a 10 Gbps SerDes lane across a 20 cm FR-4 trace with two connectors. Learn how to use S-parameter data and an eye diagram simulation to catch a failing channel before committing to fab.

Contents

The Problem With Trusting Your Layout Tool

You've routed a 10 Gbps SerDes lane — a PCIe Gen 3 or XAUI link — across a 20 cm FR-4 trace with two edge-mount SMA connectors. The DRC is green, the impedance is 100 Ω differential on paper, and the trace is straight with minimal vias. Should be fine, right?

Maybe. Maybe not. At 10 Gbps, your Nyquist frequency is 5 GHz, and FR-4 loses roughly 0.5–1 dB/cm at that frequency depending on the specific laminate grade. A 20 cm run is 10–20 dB of insertion loss before you've even touched the connectors. Add two connectors at 1–2 dB each and you're looking at a channel with 12–24 dB total loss at Nyquist — potentially enough to collapse the eye entirely.

The only way to know for sure, without spinning a board, is to simulate the eye diagram from measured S-parameters.

What You Need: A 2-Port S-Parameter File

Before opening the eye diagram tool, capture your channel response with a VNA. The file should be:

  • Format: Touchstone .s2p (2-port)
  • Frequency range: 10 MHz to at least 15 GHz (3× the data rate is a good rule)
  • Points: 1001 or more, log-spaced or linear-spaced both work
  • Port reference impedance: 50 Ω single-ended (100 Ω differential for a differential pair requires a 4-port .s4p, or a 2-port measurement of the mixed-mode S21)
Key parameters to check before you even run the simulation:
S-parameterWhat It Tells YouTypical Limit (10 Gbps)
S21 magnitude at 5 GHzInsertion loss at NyquistBetter than −15 dB
S11 magnitude at DC–5 GHzReturn loss / impedance mismatchBetter than −10 dB
Group delay variationInter-symbol interference riskLess than 50 ps pk-pk
If S21 at 5 GHz is already −18 dB or worse on the VNA, the eye simulation will confirm a closed eye — and you'll need to act before respinning.

Setting Up the Eye Diagram Tool

Upload your .s2p file to the Eye Diagram tool and configure these parameters:

ParameterValueReason
Data Rate10e9 bps (10 Gbps)Matches SerDes link rate
PRBS LengthPRBS-15Standard for BER testing; long enough to stress ISI
Samples per UI64Good time resolution without excessive compute
Input voltage swing800 mVpp differentialTypical SERDES TX swing
Rise/fall time35 ps (10–90%)Typical for a 10G TX driver
The tool convolves your channel's frequency-domain response (from the S-parameters) with the PRBS bit stream, applies the specified TX waveform, then overlays all UI periods on top of each other to build the eye.

Reading the Results: Open vs. Closed

A healthy eye at 10 Gbps should show:

Eye Opening Height150mVdiff\text{Eye Opening Height} \geq 150\,\text{mV}_{\text{diff}}
Eye Opening Width0.4UI40ps\text{Eye Opening Width} \geq 0.4\,\text{UI} \approx 40\,\text{ps}

The tool reports these numbers directly. As a rough guide:

Eye HeightEye WidthVerdict
> 200 mV> 0.5 UIPass — comfortable margin
100–200 mV0.35–0.5 UIMarginal — use equalization
< 100 mV< 0.35 UIFail — channel too lossy
A typical 20 cm FR-4 run with Isola FR408 (a better-than-commodity laminate) might produce an eye height of 180 mV and a width of 0.46 UI — marginal but passable. Switch to bog-standard FR-4 (Tg 135) and the same geometry often drops to 80 mV and 0.28 UI: a closed eye, failed link.

What To Do When the Eye Is Closed

Option 1: Reduce trace length. The simplest fix. If you can reroute to 12 cm instead of 20 cm, you recover approximately 4–8 dB of insertion loss. Re-simulate to confirm. Option 2: Switch to a lower-loss laminate. Moving from standard FR-4 to a mid-loss laminate like Isola 370HR or Panasonic Megtron 6 cuts loss at 5 GHz by 30–50%. The Controlled Impedance calculator can help you verify the new stack-up dimensions hold your 100 Ω target. Option 3: Add a CTLE or DFE equalizer. Most 10G SerDes PHYs have a continuous-time linear equalizer (CTLE) with adjustable peaking. A CTLE with 6 dB of peaking at 5 GHz can rescue channels with up to −22 dB insertion loss. Run the simulation again with the CTLE transfer function applied to see the equalized eye. Option 4: De-embed the connectors. If your VNA measurement includes fixture launches or connector pads you're not using in the final design, de-embed them. Even 1 dB of artificial loss recovery can move a marginal eye into the pass zone.

A Note on Via Stub Resonance

One failure mode that S-parameter simulation catches but layout checks miss entirely: via stub resonance. A through-hole via on a 1.6 mm board with a 0.8 mm stub resonates at approximately:

fstub=c4lstubεr3×10104×0.08×2.046.9GHzf_{\text{stub}} = \frac{c}{4 \cdot l_{\text{stub}} \cdot \sqrt{\varepsilon_r}} \approx \frac{3 \times 10^{10}}{4 \times 0.08 \times 2.0} \approx 46.9\,\text{GHz}

That's well above 5 GHz, so a standard via is fine. But a 3.2 mm stub (common if you route in the middle of a thick backplane) resonates near 12 GHz — adding a notch that chops the eye. The Via Stub Resonance calculator will flag this before you even capture the S-parameters.

Before You Send the Files to Fab

The eye diagram tool turns a gut-feel layout check into a quantitative pass/fail decision. Upload the measured .s2p, enter your link parameters, and look at two numbers: eye height and eye width. If both are in the green zone, commit. If not, you know exactly which knob to turn before you spend money on a board spin.

Run the Eye Diagram simulation

Related Articles