This article is motivated by two experiences. The first is that I attended an IEEE International Solid State Circuits Conference session on high-speed serial channels in which every single receiver was designed exclusively to operate on high loss channels. There wasn't a single presentation that considered the challenges posed by reflective discontinuities. The second experience is that we had a customer request for a tool that would synthesize a single test case that would test the eye mask compliance of a receiver. In both cases, the engineers had unrealistic expectations of eye diagrams and eye masks, and they were trying to over-simplify the task before them.
While eye diagrams and eye masks are both definitely useful, they each hide as much information as they reveal. They need to be combined in other ways to display the data, and with a clear understanding of the range of conditions to be supported. Mask compliance is no more than a beginning to what you need to know about any channel design.
Consider the following two eye diagrams (Figure 1), generated using the same realistic combination of transmit amplitude, jitter, and rise/fall time, and prepared using the CEI 28G Very Short Reach (VSR) interface specification:
Fig 1: The top eye diagram was generated from a 10.7cm channel with relatively few discontinuities while the second eye diagram was generated from a 3.6cm channel.
While the eye opening and mask compliance for the two eye diagrams look very similar, the two eye diagrams were generated for very different channels. As shown in the schematic extracts below (Fig. 2 and 3), the first eye diagram was generated from a 10.7cm channel with relatively few discontinuities while the second eye diagram was generated from a 3.6cm channel in which two 100fF shunt capacitors are separated by two symbol lengths, forming a near worst case resonant cavity.
Fig. 2: Schematic extract for 10.7cm (4.2 inch) channel
Fig 3: Schematic extract for 3.6cm (1.4 inch) reflective channel. (Resonant cavity consisting of two capacitive discontinuities separated by a 1.0cm transmission line)
The fundamental problem is that there are many different impairments that can degrade an eye diagram. And even though their root cause can be very different, requiring fundamentally different methods to control them, different impairments can look nearly the same in an eye diagram. For the example shown here, the first eye diagram is degraded by transmission line losses, resulting in a channel pulse response that is smooth, but much broader than a single data symbol. Equalization is very effective in reducing the effects of this impairment.
The second eye diagram is degraded by discontinuities in the transmission path, resulting in a channel pulse response that has a lot of ringing extending over many symbol times. The test channel design is relatively simple; however it is common for more subtle combinations of discontinuities to produce a very similar result in practical designs. It's extremely difficult to reduce this impairment through equalization, and the best solution is to identify and remove the discontinuities in the circuit board design. Some very experienced, skilled engineers have even suggested that channels should have more than some minimum insertion loss to keep the effect of discontinuities within acceptable bounds.
Figures 4 and 5 illustrate the difference in channel responses in both the time and frequency domains.
Fig. 4: Channel pulse responses for lossy and reflective channels
Fig. 5: Transfer functions for lossy and reflective channels
Here are some recommendations on what to add to eye diagrams and eye masks in order to do sound engineering of high-speed serial channels:
- In addition to looking at the eye diagrams, always look at the channel response in both the time and frequency domains. Look for channel distortion (ripples) as well as insertion loss.
In the time domain, pulse responses, both unequalized and equalized, are the most effective way to visualize the behavior at a given data rate in the time domain, although step responses such as TDR and TDT (time domain reflectometry and transmission) waveforms can also be very helpful. In the frequency domain, transfer function magnitude plots are used most often, although a polar plot or a group delay plot can sometimes be very helpful as well.
- When designing a high-speed channel or a SerDes, always consider the effects of both transmission losses and reflections. Every system has both high loss and highly reflective channels, and if you only design for one of these cases, you won't meet the needs of the system.
- When testing a design, apply many different test conditions that together cover all the different impairments you're aware of, and don't worry about the mask compliance associated with any of these tests. No single test can apply all of the impairments that may occur in practice.
- Don't depend on mask compliance for anything more than the preliminary evaluation of a design. The eye mask is a necessary lowest common denominator and may be the best the standards body could come up with. But, despite various claims, mask compliance is not a guarantee of satisfactory performance, and it is not a replacement for your engineering knowledge, insight, and judgment. Even if the design is compliant for some number of specified test conditions, your system may impose other unspecified, not unreasonable conditions which might cause the design to fail.