Digital Signal Bandwidth Vs. Frequency for NRZ and PAM-4

[@Zachariah-Peterson]

As much as I wanted to do a very long explanation on bandwidth that includes the -3dB/knee frequency, we didn't have space in this video, so I will be doing a followup to this that explains some of the big misconceptions on knee frequencies with square wave signals vs. the modulated signals discussed here. They are very different!

[@myetis1990]

Hi Zach, thank you for the great video. Looking forward to see the next video. I see that the N comes from how many bit streams can be transferred simultaneously in a modulation. So dividing full bitrate by N gives us real bitrate to handle on the transmission line. But what comes in my mind is that we usually do not specify the transmission line bandwidth to be handled directly from the data rate, for digital signals it is recommended that roughly 0.35/trise (where 0.35 may change regarding how you handle strict but anyway) , so I am confused about which one is right. is the reason that the signal forms are different ( PAM - sinusoidal signal and digital signal is square wave)?

[@Zachariah-Peterson]

@@myetis1990 The knee frequency is often misunderstoodas it relates to simple square waves, but not all high-speed signaling standards transfer data as square waves. Check out the blog in the description where I give some examples of common interfaces that use modulated bitstreams. The factor N comes from a factor 2*log(n), where n is the number of states. So for the two types being discussed, where we have 1 bit per NRZ/RZ symbol or 2 bits per PAM-4 signal, you get the factor N. If you have other PAM bitstreams you get a different number, which also depends on the encoding. For example, different Ethernet speeds use PAM-3 or PAM-5, so you don't get exactly the factor N. Here we're actually specifying the amount of bandwdith the channel needs, which includes the receiver, and it's the receiver sample rate that dictates whether it can detect a modulated signal like a PAM bitstream. The transmission line just needs to be able to transport the minimum amount of bandwidth for the receiver to do its job. For a simple square wave signal it is totally different because we are not sampling a modulated bitstream. For other signals that carry digital data there is no rise time (such as QAM) until the data is pulled from the modulated, so we can't strictly relate a rise time to the channel bandwidth.

[@myetis1990]

@@Zachariah-Peterson you are the king! thank you so much for such a great explanation :) It seems that there are a huge lot of things to learn from you.

[@Muthukumar-us8ip]

@@myetis1990 hi where can I find link about this how you handle strict. As in some places 0.5/trise has been used. So what is the trade off and benifit of that.

[@Zachariah-Peterson]

​ @Magizhini Muthukumar The 0.5 comes from the first zero in the power spectrum for a trapezoidal signal, so it is actually related to a signal's power spectrum and not the bandwidth of a channel. You can find a graph showing this on page 27 of the linked PDF document in the description. I talked about this in an Interference Technology article that should be coming out in May 2023, if you start using that formula you could underestimate or overestimage the required bandwidth. For a 112G NRZ bitstream, you would underestimate if you take rise time to be about 25% of 1 UI, but you would overestimate the same 112G bitstream with PAM-4 modulation.

[@mizunommjp]

Thank u, Zach for creating the video. I was not expected that you made the video based on my question. I understand it now. Thank you very much. My next question to you is what the title of this video says. I do not know when to use NRZ and when to use PAM4. Is PAM4 used on higher data rate? Is NRZ used on lower data rate? If the data rate is the same, PAM4 delivers faster than NRZ as it is 2 bit data trasmission whereas NRZ is only 1bit data transmission. Am I correct?

[@Zachariah-Peterson]

There is tradeoffs between these two. PAM-4 provides higher data rate but without requiring greater bandwidth in the channel. However, there will be lower noise margin between the signal levels, so intersymbol interference is more difficult to deal with. I think it's best to think about it in terms of those tradeoffs. There are some transceivers from same manufacturer where they will offer same data rate, but with either NRZ or PAM-4 on their part numbers in order to accommodate different chipsets.

[@AntonGrigorovich]

Good explanation!

[@matthiasmuller7970]

Hi Zach. Thanks for the video. I always enjoy those videos of yours, but with this one, I'm not very sure if I miss something... Like, the bandwidth if a signal in your example of a 120ghz square wave carrier with 28ghz bandwith, the signal path should actually be designed for 120+28ghz=148ghz, right? Also, to transmit a square wave, in the frequecy spectrum this is a sinx/x signal, and the bandwidth of just the square wave itself must be very high to at least transmit the first couple of harmonics. I would say, just for your example of a square wave of 120ghz, the bandwidth of the transmission line should be at least 1.2thz for the square wave... What do you think about that? Is there a reason you did not mention that, or did I miss something?

[@Muthukumar-us8ip]

I am not sure but you can think of bandpass sampling where actual signal transmission frequency is not considered only transmission bandwidth is considered for Nyqusit sampling

[@matthiasmuller7970]

@@Muthukumar-us8ip yes, but he did not talk about bandpass sampling, but about the bandwidth of a signal and the s21 and s11 of a transmition path, which should at least include the 120ght carrier, not cut off at 28ghz

[@Muthukumar-us8ip]

@@matthiasmuller7970 Okay but for digital signal the carrier will be removed right if i am not wrong. It should be a baseband or passband data maybe before DUC stage or after DDC stage where there is not even IF Freq is considered.

[@CSFitness1]

He’s not referring to a 120GHz carrier. The video mentions a 112Gbps data rate with PAM-4 signaling which has 4 amplitude levels to stuff more information into less bandwidth. So 112/4 = 28GHz, this is the information bandwidth. If instead this were 112Gbps in a 2 level amplitude signaling format like NRZ, then your information bandwidth would be 56GHz, which satisfies Nyquist sampling criteria in either case. It might seem like a no brainer to always choose PAM-4 or some higher order signaling but that requires higher SNR. As always there are trade offs to be made.

[@Zachariah-Peterson]

​ @Matthias Müller Chris has explained it correctly, the bit rate does not equal the frequency of some carrier.

[@kecsrobi6854]

What would be a safe margin for over design? in this case you mentioned 28GHz but would it be good enough to design for 30GHz? or do I need to take into account a larger margin or error ? asking since this way I may get a bit more or a leway to mess up a bit. Also if this trace will be exposed how will surface finishes generally affect it? like ENIG ?

[@CSFitness1]

Answer will always be it depends. Surface finish will have an impact on channel quality as you approach higher frequencies, this is the value of simulating channels using field solvers like Hyperlynx, ADS, etc.

[@Zachariah-Peterson]

I love this kind of question because at the end of the day there is no "rule of thumb" for the margin that should be applied. At the end of the day, the channel's bandwidth is not the only metric that matters, what matters is that you sit within the S-parameter mask for S11, S21, and crosstalk. There is also bit error rate. In total, these metrics are used for channel compliance.

[@ravirajsingh4785]

Hi Zach, this session was awesome. In the automotive industry isolation is required, Can you make one video related to isolation(digital and voltage).

[@Zachariah-Peterson]

Sure I would be happy to, it's usually with isolated components or with optocouplers, but I'll see if I can put together a good tutorial.

[@aopfin]

Hi, As a new user it has been a bit difficult to get into the Altium Signal Integrity tool as the documentation on Altium website is out-dated and there are no tutorials or webinars available. Would it be possible to do a tutorial of the current Signal Integrity tool and how to correctly set up a SI simulation with IBIS models (for example CPU+RAM or some other bus)? It would also be nice if some simulation pitfalls were covered (limitations, risk and types of user errors etc).

[@Zachariah-Peterson]

This is a great topic, sure I will set something up with IBIS models. Is there any specific component that would be helpful to look at? Or do you have a specific model you are trying to work with?

[@aopfin]

@@Zachariah-Peterson IMX6Q with Micron DDR3 (specific model doesn't matter, just the setup process). There is the IMX6rex open source project which uses the same components and is available for download. It would be nice to know how Altium SI calculates stuff like via transitions.

[@Zachariah-Peterson]

​ @aopfin Unfortunately Altium does not calculate via transitions, although I have complained to product people that they need to integrate the via tool from Simbeor. The layer stack manager currently uses the Simbeor SDK to do impedance calculations, so they could definitely create a via impedance tool, something that includes auto placement of stitching vias as well would be very nice.