Yankees versus Red Sox or Real Madrid versus Barcelona or NRZ versus PAM4

Even though I would love to write about all those different rivalries, the focus of the post today will be on NRZ versus PAM4. We will talk about them in a minute, but in order to talk about it, first we have to define what a digital sequence is. Suppose you want to talk to someone over your phone. In the digital era, you don’t send your voice as it is, but a quantified version of it. Please take a look on the picture below:


As you can see, your original voice is converted to a signal with different levels (in our example, eight levels). Moreover, each level is associated with a unique sequence of three bits. This is done in such a way that the normal human ear cannot tell the difference. Your voice can now be represented by a (very large) sequence of bits, with every three bits representing one level. Your phone will do this process and send the bit sequence. On the other side, whoever you are talking to will undo this process (or better yet, her phone will): receive the sequence of bits and regenerate your voice signal. If you want to learn more, please go to http://www.molex.com/training/ee_signal/course.html

This quantization process is the same if you want to transmit voice, video or data. Now imagine a lot of people wanting to talk, watch videos, and send data, all at the same time. You will need a system that can handle a very large number of bits. Within Molex, anything beyond 1 Gbps (more than a billion bits transmitted in a second), is called high-speed. We are now developing the connectors that will handle 25 to 40 Gbps. But there are challenges here, and some creative solutions have been proposed. Today we will talk about one of them: PAM4.

The most obvious way to send those bits is to simply generate a voltage signal that has values 0 and 1. This is called non-return to zero (NRZ); even though it does return to zero (don’t we love engineers?). But there is a problem there. Ideally, the voltage should go from 0 to 1 and back within the time a single bit is allowed to exist. However, a real circuit will generate a voltage that is far from ideal. Please take a look on the picture below.


As you can see, after travelling through the channel, your bit will be attenuated and distorted, interfering with the next bit. Now suppose you want to send even more bits (our customers always want to send more bits!). The more bits you want to send, the faster it will have to go up and down, and the interference will affect more and more bits, making it almost impossible for the system to tell if what it sees is really the number “1” or the interference left from the previous bit.


So what to do? What if we could send two bits packed in a single symbol? As a matter of fact, we can do that. Rather than sending and reading only 0s and 1s, we will now use 0s, 1/3s, 2/3s and 1s. A 0 corresponds to 00, a 1/3 corresponds to 01, a 2/3 corresponds to 11 and a 1 corresponds to 10. The picture below shows how this process is done.


Notice that the original sequence had 22 bits to be transmitted, while the PAM4 sequence has only 11 symbols (0, 1/3, 2/3 and 1), but the same information is available in each case. With this simple trick, we can send twice as many bits at the same speed. This is called 4-level pulse amplitude modulation (PAM4). So all of our problems are solved, right? If you want to send more bits, just pack them together in a symbol and send the symbol instead, correct? Not quite so…

The limitation of PAM4 is two-folded: the first problem is that we reduced the amplitude of our symbol by a factor of 3. So, if before the attenuation of a single bit was already a problem, now the problem is even worse: now we cannot afford much more attenuation, as we are giving away 2/3 of our original amplitude to go to PAM4.

The second problem is that most of the systems in use today are based on NRZ signal, which is quite simple and cheap. Therefore, implementing a circuit that will encode the bits into a symbol and then decode them back will cost money.  Ultimately, the performance figure that really matters for our customers is the cost per bit, so paying more to send more bit is not always an attractive solution.

All in all, PAM4 could be a solution if there is no way out: you cannot possibly transmit more bits over your channel and you can afford some extra attenuation. It requires (roughly) half the bandwidth of a NRZ signal, as it can send twice as many bits, but it attenuates the signal by a factor of 3. And it will also cost more money. So the best solution will depend on the problem at hand. No silver bullet here. And the rivalry will go on.