|PAM4 doubles the number of bits in serial data transmissions|
by increasing the number of levels of pulse-amplitude modulation,
but does so at the cost of noise susceptibility
For quite some time, NRZ-type encoding has been the mainstay modulation scheme for data transmission. In an NRZ scenario, we take a binary pattern, say 0011010, and encode that into a series of fixed voltage levels, with the lower voltage being a zero and the higher voltage being a one (see data stream M in the figure). We'll assume a given bit rate of, say, 28 Gb/s.
If we look at that NRZ signal as an eye diagram, it will have a bit period, T, and amplitude, A. The required bandwidth for this signal is related to the bit period (1/T). The faster the bit rate, the shorter the bit period and the higher the bandwidth.
There is also a signal-to-noise (SNR) ratio requirement, which is related to the amplitude, A. The smaller the eye diagram becomes vertically, the more difficult it becomes to maintain a SNR that allows us to interpret the signal at the receiving end of the link.
Fundamentally, what we'd like to do is to double the number of bits we send from point A to point B. One way to achieve this goal is by adding a second lane or channel. In this channel, we might like to send a different bit pattern, say 0101100 (see data stream L in the figure). But there's a downside to this approach, which is that we now need two transmitters, two receivers, and two channels. We might not have the luxury of the additional real estate or power consumption, so we look for another solution.
What else might we do to double the bit rate? One approach is to serialize the two bit streams. Instead of two 28-Gb/s channels, we create one 56-Gb/s channel. As a result, in the same period in which we had one bit transmitted at 28 Gb/s, we now have two bits transmitted at 56 Gb/s. That would look like the bit stream ML in the figure.
The eye diagram for signal ML shows that the amplitude is still the same as it was for signals M and L, but the period is now T/2. If we flip that number upside down, we get the bandwidth, 2/T. We retain the SNR requirement related to A, but the required bandwidth for the signal has doubled. So it's good news and bad news on SNR and bandwidth, respectively.
We need a way to double the bit rate in the channel without doubling the required bandwidth, and that's where PAM4 enters the picture. PAM4 takes the L (Least Significant Bit) signal, divides it in half, and adds it to the M (Most Significant Bit) signal. The result is four signal levels instead of two, with each signal level corresponding to a two-bit symbol.
The PAM4 signal looks like trace M+L/2 in the figure. At the lowest level is 00, followed by 01, 10, and 11, respectively. PAM4 indicates pulse-amplitude modulation, with the "4" indicating four levels of pulse modulation.
An eye diagram for a PAM4 signal is unusual, with three eye openings and four levels stacked vertically as shown in the figure. The bit period (or symbol period) is T. However, the opening of each of these three eyes is A/3. For bandwidth requirement, we roll back to 1/T. Thus, this signal, which moves 56 Gb/s, does so using the same amount of bandwidth as did the ML signal that moved 28 Gb/s. But with the SNR related to A/3, we find that our M+L/2 signal is three times more susceptible to noise.
We have, in effect, traded off SNR for bandwidth. Many serial links are bandwidth-constrained, as it's difficult to move much more than 28 Gb/s over any length of copper. But when you have some SNR headroom, it may well pay off to consider a PAM4 modulation scheme.
Now that we've covered the basics of PAM4, we'll look next time at real-world application scenarios and consider the test requirements related to each one.