For essentially the entire history of pluggable optical transceivers the bit rate and the baud rate were identical. Since pre-400G transceivers used Non-Return to Zero (NRZ) signaling, a form of simple 2-level Pulse Amplitude Modulation (PAM-2), only a single bit of information was encoded in each transmitted symbol. Therefore, the symbol rate, also known as baud rate, was the same as the bit rate. NRZ signaling is very simple and worked well up to 25Gbps. NRZ is limited to symbol rates below about 32Gbaud. That limit was fine for 100G QSFP28 optical transceivers with their 25G channels. However, 400G QSFP-DD transceivers, with their 50G (actually 56G with FEC overhead) channels, require a more exotic modulation scheme. Standards for 400G have settled on PAM-4 signaling. This blog will describe the differences and benefits of PAM-4 versus NRZ for 56G channel transmission.
NRZ vs. PAM-4: What’s the difference?
NRZ signaling encodes a single bit of data in each transmitted symbol. As mentioned above, NRZ can also be referred to as PAM-2. Each bit is transmitted as one of two amplitude levels, 0 or 1 (see figure 1). The bit rate of an NRZ transmission is, therefore, equal to its symbol/baud rate.
In PAM-4 signaling, each symbol is one of four distinct amplitude levels (shown as 00, 01, 11, and 10 in Figure 2). Since each amplitude level represents two bits, the baud rate is half the bitrate. Therefore, the 56Gbps channels used in 400G optical transceivers requires a symbol rate of only 28Gbaud.
Ok, So What’s the Downside of PAM-4?
The best way to understand the downside, or additional complexity, of PAM-4 signaling versus NRZ is to compare what are known as ‘eye diagrams’ for each, shown in Figure 3. These diagrams are screenshots from oscilloscopes synchronized to the 56G signals. In the figure, the top eye diagram (NRZ) clearly shows a signal with two distinct amplitudes as well as the transitions between these levels. The “eye” is the open space created in the middle of these transitions. The larger the eye the more clearly delineated are the two possible levels in this signaling scheme.
The bottom diagram in Figure 3 shows the three ‘eyes’ separating the 4 distinct amplitude levels of a PAM-4 signal. As expected, the eyes are approximately 1/3 the height of the NRZ eye. This results in a loss of more than 9.5dB in Signal to Noise Ratio (SNR) of PAM-4 versus an NRZ. This smaller eye opening substantially reduces tolerance to impairments such as reflection and crosstalk. Therefore, PAM-4 has a much higher inherent Bit Error Rate (BER) than an NRZ signal of equivalent baud rate.
To overcome the high BER issue much more sophisticated signal processing is required. Each of the three PAM-4 eyes is slightly different, requiring distinct signal crossing, data and error detectors and each eye must be independently equalized. Finally, Forward Error Correction (FEC) is necessary to recover errored data to achieve an acceptable channel Bit Error Rate. We will discuss the role of Forward Error Correction in another article.
Conclusion
To enable processing, storing and access to ever increasing data-intensive activities, hyperscale data centers are rapidly moving beyond 100G to 400G interconnections. These 400G Ethernet links are based on PAM-4 signaling for their 56G channels. This is a major step away from the tried-and-true NRZ (PAM-2) signaling that has been the standard module scheme for pluggable optical transceivers to date. While PAM-4 delivers twice the bit rate (bits per second, bps) for a given baud rate, this comes with greater sensitivity to channel impairments. The good news is, advanced signal processing ASICs and Forward Error Correction (FEC) are built into a new generation of QSFP-DD optical transceivers. Therefore, end users can deploy these new 400G devices in essentially the same plug-and-play fashion as previous generations.