The next generation's modulation: PAM-4, NRZ, or ENRZ?
As people use more and more apps on their phones, tablets, computers, and IoT devices, the network needed to deliver the data for those apps constantly needs upgrades. By now, 28 Gbps serial channels that run on between ICs on a board, between boards over backplanes, and over fiber is well on its way to deployment. Now, it's time to define 56 Gbps channels.
Every speed increase brings new problems along for the ride. At OFC 2014, Nathan Tracy of TE Connectivity moderated a panel called OIF Physical Link Layer Session "56 Gbps Serial—Why, What, When" where three engineers made their cases for which form of modulation will become the 56 Gbps standard. The panelists and their modulations were:
- Ed Frlan, Semtech: PAM-4
- Brian Holden, Kandou: ENRZ
- Tom Palkert, Luxtera: NRZ
Frlan opened with a discussion of why NRZ won’t work at 56 Gbps and why PAM-4 is the way to go. Figure 1 shows how PAM-4 uses four distinct amplitude levels, each of which represents a combination of two bits (00, 01, 11, 10) where each level is clocked on a rising or falling edge of a clock signal. The left side of the oscilloscope screen is the signal itself with the right side showing how the four levels combine to form three levels of eye diagrams.
Figure 1 PAM-4 uses four distinct voltages to represent two bits (left), which produces three eye diagrams (right).
PAM-4 is gaining traction in 28 Gbps links, shown both at OFC 2014 and at DesignCon 2014as an alternative to NRZ modulation. But, Frlan argued that channel losses for 56 Gbps NRZ are too great to support CEI-56G-VSR (very short reach) links, citing a 14-dB channel loss at 14 GHz (Figure 2). He also claimed that a 100 Gbps, 4x28 Gbps link will have too much crosstalk at 56 Gbps NRZ to make it feasible.
Figure 2. According to Ed Frlan, NRZ signals will have too much loss 10 dB at 14 GHz in a VSR channel.
Frlan showed that for a error rate of 1E-6, each eye opening is about 70 mV for each of the three eye openings in a PAM-4 signal. Of course, such an error rate is unacceptable. The eye height shrinks to about 40 mV when the error-rate drops to an acceptable 1E-15. Figure 3 shows the difference in a voltage bathtub curve (left). The right side shows a bathtub curve for the eye width.
Figure 3. Eye openings shrink both vertically (70 mV to 40 mV) and horizontally (0.28 UI to 0.16 UI) when error rates drop from 1E-6 to 1E-15.