Issue link: https://iconnect007.uberflip.com/i/1014812
AUGUST 2018 I DESIGN007 MAGAZINE 67 showed variations consistent with the low- cost PCB interconnects. This correlation was sufficient to reliably predict the eye diagram for the 28 Gbps NRZ signal as seen in Figure 3 (15 ps rise and fall time, PRBS-32). A 5.5% difference in the eye heights and just 1.5% in eye widths is not bad, considering the low-cost manufacturing very similar to the production boards. Now, let's try to predict 56 Gbps PAM-4 with the same 15 ps rise and fall time and PRBS-32. Notice in Figure 4 that the relatively small amplitude error in the case of NRZ is now about a 25% error! The eye width is not pre- dicted accurately either. What is the problem? Obviously, we have smaller eye openings for three eyes and, naturally, the relative error increased. In Figure 5, let's look at the spec- tra of the signals used to generate these eye diagrams (computed with DFFT for the finite sequence of bits and symbols). The NRZ signal in this case has about twice the power compared to the PAM-4 signal with the double data rate and the same rise/fall time and the same amplitude. The ratio of the power in the second lobe to the first one is about the same for both signal types. The bit time of the 28 Gbps NRZ signal is equal to the symbol time of the 56 Gbps PAM-4 signal. The accuracy of the analysis for both signals depends on the accuracy of the model above 28 GHz, that is where the second lobe of the spec- trum is located. The model in this case does not correlate with the measurements above about 30 GHz due to the loss of the localiza- tion properties by the vias. This inaccuracy did not prevent acceptable correlation for the 28 Figure 4: Moving from 28 Gbps to 56 Gbps raises the NRZ error to about 25%. Figure 5: The spectra of the signals used to generate these eye diagrams.