The Channel: Loss, Reflections and Crosstalk
Measure a channel the way its datasheet does: insertion loss, return loss and crosstalk, and the loss budget a link must fit.
You Will Be Able To
- Read insertion loss from a frequency response and relate it to length.
- Predict the reflection from an impedance step and convert it to return loss.
- Explain near-end and far-end crosstalk.
- Check a channel against a loss budget at the Nyquist frequency.
- Lesson
- P10.8
- Time
- 30 min
- Runs on
- ngspice
- Check
- 5 questions
The same 10 Gb/s link works on a short board and fails on one twice as long. Doubling the bench channel took its loss at 5 GHz from 3.7 to 13.0 dB.
- I/O
- Channel / handshake
- Measure
- Data
- Send a wave
A sine wave of one frequency goes in at port 1.
- S21: what gets through
The ratio of the wave out of port 2 to the wave in. In dB it is negative; its size is the insertion loss.
- S11: what comes back
The ratio of the wave reflected at port 1 to the wave in. Its size in dB is the return loss: larger is better.
- Differential names
For a differential pair the names gain two letters: SDD21 for the differential insertion loss, SDD11 or SDD22 for return loss.
Read the diagram as text
- Port 1: transmitter side
- Channel: trace, connector, cable
- Port 2: receiver side
- S11: return loss
- S21: insertion loss
- Port 1 → Channel: incident
- Channel → Port 2
- Channel → S11: reflected
- Port 2 → S21: transmitted
USB4 writes its transmitter return-loss mask as SDD22 (§3.1.3.1.1.2). S12 and S22 are the same ratios seen from the other end.
Insertion Loss
Insertion loss is how much of a signal a channel loses between its two ends, at each frequency. Traces lose more at high frequencies, so the loss curve falls with frequency.
- k = 0.3
- k = 0.6
- k = 1
- k = 1.5
Read the plot as a table
| Frequency | k = 0.3 | k = 0.6 | k = 1 | k = 1.5 |
|---|---|---|---|---|
| 10 MHz | -0.000004951 dB | -0.00002294 dB | -0.00007663 dB | -0.0002132 dB |
| 100 MHz | -0.0004951 dB | -0.002294 dB | -0.007657 dB | -0.02127 dB |
| 1 GHz | -0.04925 dB | -0.2238 dB | -0.7092 dB | -1.764 dB |
| 10 GHz | -3.32 dB | -8.282 dB | -14.53 dB | -22.8 dB |
ngspice AC runs. Loss at 5 GHz for k = 0.3, 0.6, 1, 1.5: 1.1, 3.7, 7.8, 13.0 dB.
The number that matters is the loss at the Nyquist frequency, half the bit rate: 5 GHz for 10 Gb/s. A bit pattern of 1010 is a 5 GHz square wave, the fastest thing the link sends.
On the bench the loss grows faster than the length. At 5 GHz, k = 0.6, 1 and 1.5 lose 3.7, 7.8 and 13.0 dB.
Reflections and Return Loss
A line has a characteristic impedance Z0. A wave travels along it unchanged until the impedance changes. There, part of it reflects.
The reflected fraction is Γ = (ZL − Z0) / (ZL + Z0). A matched end reflects nothing; an open end reflects everything; a short reflects everything with its sign flipped.
- end matched
- end twice Z0
- end half Z0
- end open
Read the plot as a table
| Time | end matched | end twice Z0 | end half Z0 | end open |
|---|---|---|---|---|
| 0 s | 0 V | 0 V | 0 V | 0 V |
| 200 ps | 500 mV | 500 mV | 500 mV | 500 mV |
| 400 ps | 500 mV | 500 mV | 500 mV | 500 mV |
| 600 ps | 500 mV | 666.7 mV | 333.3 mV | 1 V |
| 800 ps | 500 mV | 666.7 mV | 333.3 mV | 1 V |
| 1 ns | 500 mV | 666.7 mV | 333.3 mV | 1 V |
ngspice transient of a lossless line. The level after the reflection tells the load: 0.5 V matched, higher for a larger load, lower for a smaller one.
The bench read the sending end 800 ps after the step. Matched, twice Z0, half Z0 and open gave 0.500, 0.667, 0.333, 1.000 V: 0.5 × (1 + Γ).
Return loss is the reflection in dB, −20 log10 |Γ|. A load of twice Z0 reflects a third: 9.5 dB of return loss. Larger is better.
USB 3.2 limits the integrated return loss of a linear re-driver to −17 dB, against an 85 Ω reference (§E.6.4.5). USB4 gives its transmitter a return-loss mask, SDD22, from −8.5 dB at low frequency (§3.1.3.1.1.2).
A reflection does not vanish. It travels back and reflects again at the source if that is not matched. It reaches the receiver late, as a ghost of an earlier bit.
Crosstalk
Two traces side by side are coupled by the capacitance between them. An edge on one, the aggressor, puts a small pulse on the other, the victim.
- aggressor drive (left axis)
- victim near end (right axis)
- victim far end (right axis)
Read the plot as a table
| Time | aggressor drive | victim near end | victim far end |
|---|---|---|---|
| 0 s | 0 V | 0 V | 0 V |
| 200 ps | 1 V | 12.88 mV | 14.62 mV |
| 400 ps | 1 V | 2.254 mV | 2.559 mV |
| 600 ps | 1 V | 262.3 µV | 297.7 µV |
| 800 ps | 1 V | 27.09 µV | 30.76 µV |
| 1 ns | 1 V | 2.666 µV | 3.027 µV |
ngspice transient. The victim's near end peaks at 16.1 mV and its far end at 17.8 mV.
Near-end crosstalk (NEXT) appears at the victim end next to the aggressor's driver. Far-end crosstalk (FEXT) appears at the end next to the aggressor's receiver.
On the bench a 1 V step gave 16.1 mV of NEXT and 17.8 mV of FEXT: -35.8 and -35.0 dB.
USB 3.2 limits the integrated NEXT and FEXT of a re-driver to −35 dB between any two pairs (§E.6.4.6). It notes that the re-driver's gain amplifies the crosstalk it lets through.
Crosstalk is noise that tracks a neighbour's data. An equalizer that boosts high frequencies boosts it too.
The Loss Budget
A link is designed to a loss budget: the largest insertion loss at Nyquist that its equalizers can make up.
| Budget | Loss at Nyquist | Source |
|---|---|---|
| USB 3.2 host or device (informative) | 8.5 dB | §E.6.4.8 |
| USB 3.2 end to end (informative) | 23 dB | §E.6.4.8 |
| USB4 Gen 2 receiver test, maximum-loss case | 18 dB at 5 GHz | §3.1.4.2 |
| AMD transceiver wizard: DFE mode above | 14 dB | PG182 |
| Bench channel at k = 1.5 | 13.0 dB at 5 GHz | ngspice run above |
The bench rows are bench values. USB 3.2 calls its budgets informative.
USB 3.2 shows a channel that loses about 33 dB, well past its 23 dB budget (§E.6.4.8). Even the highest re-driver gain cannot rescue it.
P10.9's passive CTLE lifted 5 GHz by about 9.9 dB against DC. That covers the long bench channel's 7.8 dB, but not the 13.0 dB of k = 1.5.
USB 3.2 adds that insertion loss is not the only limit: crosstalk and reflections also eat the margin (§E.6.4.8).
Try It
Four copies of one line, each with a different end. Each meas line reads the sending end after the reflection has come back.
The matched line reads 0.5 V: the source and the line form a 2 : 1 divider, and nothing returns.
Change one end to 5 Ω and predict the level before you run: Γ = −45 / 55, so about 0.09 V.
Run the netlist to draw its plots here.
back0 ≈ 0.5 back1 ≈ 0.6666667 back2 ≈ 0.3333333 back3 ≈ 1Check Yourself
Common Mistakes
Wrong Quote a channel's loss without a frequency.
Loss grows with frequency; the long bench channel loses 7.8 dB at 5 GHz and almost nothing at 10 MHz.
Fix Quote insertion loss at the Nyquist frequency of the bit rate, and compare it with the budget there.
Wrong Call a return loss of 3 dB good because the number is small.
Return loss is the reflection's size below the incident wave; a small number means a large reflection.
Fix Look for large return loss, such as the 17 dB USB 3.2 asks of a re-driver.
Wrong Terminate an 85 Ω differential pair with 50 Ω per pair.
Any step in impedance reflects; 50 Ω against 85 Ω reflects about a quarter of the wave.
Fix Match the termination to the line's impedance, differential to differential.
Wrong Treat crosstalk as random noise.
It follows the aggressor's data and edges, so it is bounded, deterministic and correlated with a neighbour.
Fix Budget it as deterministic jitter and noise, and remember that equalizer gain amplifies it.
Remember
- Insertion loss (S21) is what a channel loses end to end; quote it at the Nyquist frequency.
- An impedance step reflects Γ = (ZL − Z0) / (ZL + Z0); return loss is that reflection in dB.
- Crosstalk couples a neighbour's edges onto a victim, at the near end and the far end.
- A link is designed to a loss budget at Nyquist that its equalizers can make up.