What’s effective return loss, anyway? (Part 1)

Article By : Ransom Stephens

We used to call it ringing, and we fixed it by limiting return loss. Now, we’ve invented effective return loss, which made it both complicated and flexible.

Remember ringing? The oscillations that infect signals and are caused by reflections? Back when you could assume that signals traveled from chip-to-chip instantaneously, we either fixed the problem by matching impedances, not such a big deal at MHz, or waited for them to settle down (Figure 1).

Reflections produce ringing in rising edges.Figure 1. Reflections produce ringing in rising edges.

I don’t remember those days, either. The speed of light is decidedly finite. The difference between then and now is that the time it takes a signal to travel from transmitter to receiver, reflect from the receiver back to the transmitter, and then reflect a third time from transmitter to receiver is much less than a symbol unit interval (UI)—whether it’s a non-return-to-zero (NRZ) or a four-level pulse amplitude modulation (PAM4) signal.

In gEEk terms, the path from transmitter to receiver is a transmission line, as fundamental a network element as a capacitor, inductor, or resistor. The UI of a 56 Gbaud signal is less than 20 ps, which spans about 3 mm in typical PCB; any channel longer than several a couple of centimeters would qualify as a transmission line for this signal.

Think of it like this: Impedance mismatches at the pins of the transmitter and receiver plus connectors, vias, and other discontinuities between them cause reflections. If the distance between the transmitter and receiver is an inch, then the group delay between them is about 8.5 UI for a 56 Gbaud signal. The reflection of a symbol at the receiver travels 8.5 UI back to the transmitter, experiences a secondary reflection and travels another 8.5 UI to the receiver. Since the round trip takes 17 UI, that reflection degrades the symbol that was transmitted 17 UI after the original symbol. A 17-tap decision feedback equalization (DFE) is perfectly suited to tidy up these reflections before the symbol enters the decoder/slicer.

Due to insertion loss and the fact that none of the reflections are perfect, subsequent reflections have ever smaller amplitudes.

RL(f) (return loss as a function of frequency) and IL(f) (insertion loss) are given by the differential scattering parameter Sdd22 and Sdd21, respectively. Sdd22 measures the total reflected signal energy. S-parameter masks have been used for years to specify the maximum allowed RL(f) and IL(f), but they don’t account for the effects of equalization (Figure 2).

Masks for a typical PAM4 28 Gbaud applicationFigure 2. (a) RL(f) and (b) IL(f) masks for a typical PAM4 28 Gbaud application. Courtesy of Ransom’s Notes.

Enter effective return loss (ERL), a quantity introduced in 802.3cd (50/100/200/400 Gigabit Ethernet) by Rich Mellitz, a Distinguished Engineer at Samtec. ERL incorporates return loss with the effects of equalization—especially DFE—as well as transmitter noise and receiver frequency response into a signal-to-noise-like figure of merit in a way that is similar to channel operating margin (COM).

Like COM, ERL does two things: (1) it provides a flexible design parameter space in which engineers can optimize their designs for the system as a whole, allowing for different elements of the design to accommodate different signal impairments while assuring that compliant components will be able to interoperate. And, (2) it takes what had been a simple, easy-to-understand measurement with a cut-and-dry requirement and turns it into a figure of merit that’s so complicated that you have to read part 2 of this article to understand what it is. But don’t worry, part 2 will post in March, at which time the link will take you there.

Ransom Stephens is a technologist, science writer, novelist, and Raiders fan.

Leave a comment