[00:00:12] Hello everyone and welcome back to the Cisco Optics Podcast where we talk about pluggable optics for networks. In the pluggable optics industry, every next generation is led by countless presentations and proposals at standards bodies meetings along with lots of lab development and testing behind the scenes. Right now, the industry has its sights set on 3.2 terabit transceivers with 400 gigabit per electrical lane. All those meetings and all that testing are happening as we speak.
[00:00:39] Today, I am joined by John Calvin, Strategic Planner and Datacom Technology Lead at Keysight. John has been bridging the measurement science gaps of emerging telecom and datacom development efforts for 20 years. He has chaired multiple physical layer interoperability working groups and is currently serving as contributing member to IEEE 802.3, OIF-CEI, InfiniBand and PCIE development efforts.
[00:01:04] John holds a BSEE from Washington State University and his graduate level studies are in signal processing from Stanford University. In episode 67, we continue our conversation with John. We get into how equalization, forward error correction and increasingly sophisticated signaling keep Ethernet scaling. And now join me as I talk with John Calvin.
[00:01:25] How quite a wide range of interoperability in their career, we learn about 10 gigabit per second.
[00:02:03] in IEEE for a number of years. They've been doing PAM5 for many years down there. But the idea was you're dealing with CAT5 and CAT6 cables that had physical bandwidth limitations on less than a gigahertz, right? So you had to, the only place you could go was to go to higher order modulation, stack more bits up per unit interval, and they did PAM5. It was commonplace back then. Oh, okay.
[00:02:30] So, you know, higher order modulation is not a new thing. But the thing that really has been a key enabler is you think about what we had for standard sort of data center interconnects. SFP module back in 2010 would be a 10 gigabit per second. It'd have 10 gig by four lanes going into this thing,
[00:02:59] and it'd be 10 gig NRZ, right? So a QSFP. A QSFP module or something along those lines. And at that time, 10 gigabits per second NRZ was blisteringly fast, right? Yeah. I remember when we were building BIRT's for those speeds back then, that was actually a big deal. You know, we had a Vantes D3186 BIRT's. It cost half a million dollars. They were 10 gig BIRTs.
[00:03:28] And they were the essential tools to make those interfaces work and to be tested. So that happens to be when I went, the time that I got into transceivers, 40 gig was the latest thing at that time. So the big story was like, wow, we just went from 10 gig to 40 gig by just having four lanes in a single form factor, right? And they're playing this game where you can go multiple lanes,
[00:03:58] you can aggregate this different ways and build on the success of a single lane rate. So 10 gig was a very solid technology, and it got scaled up. There's even like SR10, which was 10 lanes wide of multimode at 10 gigabits per second. It was a very successful platform to build up from. So what was the next step in the path toward 400 gig? What was the next stepping stone?
[00:04:24] I think if you look at the architectures of the equalization that you find inside these FIs, now, and SIRDEs, transmitters, receivers, at 10 gig, it was actually an openized spec, right? So we had a CTLE as a form of equalization. It's a very simplistic sort of RC network equalizer.
[00:04:49] And that was the equalization mechanism of choice back then. You didn't have advanced feedforward equalizers or decision feedback equalizers. You never had MLSE back then. But as we have gone from standard to standard, so we go, let's just say, we went from having a FIR filter and a transmitter to having a CTLE and a receiver
[00:05:17] to having a CTLE and a FIR filter and a receiver to having a CTLE, a FIR filter, and a DFE in the receiver. So the receive chain for the equalizers has simply gotten larger and larger. More tools, more assets, more technology has been applied to compensate for the losses of these channels. The channels get worse, the compensation gets better, and they balance each other out.
[00:05:46] So right now we're at a point where in the DJ specification at 200 gigabits per second, we have a transmit FIR filter, a receive CTLE, a receive FIR filter, a receive DFE, and an optional MLSE. There are five equalizers in the signal path. Most of them in the receive side? Well, the majority is on the receive side, yeah.
[00:06:15] Are these all dynamically tuned or with a feedback group? They are dynamically tuned. There is a concept known as interlink training. So the links are actually designed to communicate. There's a communication protocol that's in-band that allows it to say, all right, step up your coefficients or modify your transmitter to give me more energy on the main cursor or something like that.
[00:06:44] And it can basically auto-tune in the link in order to improve the BER. And then it goes into operational mode. And when it does go into operational mode, does the tuning stop, or does it continue to monitor? Yeah, it stops at that point. There are points in time where you may have a link failure and it automatically retuned. Or you can also monitor your BER to see if it's creeping up. Because, you know, things change in a data center. Things get hot.
[00:07:14] There's movement of cables and the cable chase and stuff like that. And occasionally you have to retune your link. But basically it's on power up. And it powers up and it gets into a state where it's ready to roll. It's ready to transmit. And then it's all tuned up and it can broadcast bits. That's where it lights up the line. And we'll send 200-gig traffic today and 400-gig traffic in a couple of years.
[00:07:44] Like you said, it's pretty exciting. Well, you know, you think about the evolution of the equalizer structure. There may be some points of limited return here, right? Because every time you put a new equalizer structure in there, it's adding latency. It's gates, right? It's burning power. So, I mean, they're fighting this force where they want to have, you know,
[00:08:12] one and a half meter link performance. And you want to be able to connect to transceivers and modules outside of your switch. But at the same time, you don't want to pay that penalty of power consumption and all the other complexities that come with it. And that's what's driving these scales into co-packaged copper, co-packaged optics. It will happen. It's only a matter of time.
[00:08:40] But like you said, you can't have a multi-vendor strategy. Yeah. Until you could also get this sort of thing standardized. Once you walk away from the pluggables, right? You could standardize a co-packaged copper interface and say, this is your footprint. Everyone can build this footprint. And it's a vendor free-for-all who wants to build a transceiver to put on these co-packaged copper landing points around the 30s. And that idea is being led by various MSAs right now.
[00:09:10] There are MSAs actively pushing the standardization process for this interconnect. And the reality is every connector vendor out there wants their connector vendor to be the standard, right? So that's why there's not a lot of agreement right now because everyone thinks their connector is going to be it. So these packages will be placed onto the PC board with a socket connector or would they be soldered on?
[00:09:36] There's a micro-BGA and there's a couple different ways of doing this. And again, it's not a standard interface. But the ways that I've seen it, if you have a small square connector about the square, about the size of a dime, that kind of size right there. And it's going to have 64 differential pairs on it and an elastomeric bonding agent that allows you to mechanically press this thing onto a micro-BGA. I've seen it work.
[00:10:04] It's actually, the performance is actually remarkably good. It's not the kind of thing where you can actually make and break this thing many, many times. They haven't quite perfected that technology yet. But it's the kind of thing where if you're trying to configure a switch and you want to say, all right, I want these ports to be coherent. I want these ones to be DR8. And I want these ones to be multimode. You'd have the ability to configure it at your production time when you're setting up in a fab and you're trying to make this thing work and configure it.
[00:10:34] Okay. So you can't necessarily change it in the field, but you can at least have some flexibility at assembly time. You know, and really the goal is to have this thing feel updatable. One of the problems here is that, you know, optics is really where people want to go. They want to go to a Vixxel-based design or go to a thin-film lithium-naubate-based modulator.
[00:11:02] But the problem is lasers and optics still have a significant infant mortality problem. And for reasons that are really hard to nail down from a material science standpoint, optical interfaces are not that reliable.
[00:11:21] And the idea of, you know, taking your module transceivers and putting them down around your co-packaged copper setup and sliding it into a rack and never touching it again is a fallacy that hasn't been really proven out. And that's one of the attractive things about the whole modular interface, which is when your transceiver gets a bad port on it, well, you can identify it with software. Let's say your transceiver is bad.
[00:11:48] You can query through CMOS and find out which one is bad, pull it out and swap it. And you have a field technician in the back of a rack for 10 minutes and you're done. And this way you've got to pull out the whole tray. You have to have people in a clean setting to be able to pull these BGAs apart and reassemble them. And that's imperfect science at best. So that's really the challenge right now that people are trying to overcome.
[00:12:14] If they could make the optics really, really reliable, solve that problem, then they would never have a scenario where they'd have to open up the chase and go into a switch chip and rechange or reconfigure its optical or copper interconnects. And that's one of the things why people lean on copper right now is that, you know, you could tow a truck with a copper cable and it's still plugging in your data center and it'll work just fine.
[00:12:43] You could slam the door on your copper cable. It's going to work just fine. The reliability of copper cable, especially passive copper, is just stunningly good. It doesn't wear out. It doesn't break. And that's one of the reasons why copper is still in the game. It's, I mean, as we go to higher data rates, is there a point where copper just throws up its hands and says, okay, I can't do a meter anymore?
[00:13:07] God, you know, I used to keep a data sheet as an article from the... That was the fourth part of my conversation with John Calvin. Next time we'll get into the limits of copper and emerging alternatives such as plastic waveguides and hollow core fiber.
[00:13:38] Please follow the Cisco Podcast Network on Spotify or Apple Podcasts. Better yet, leave a review on Apple Podcasts. Remember, you're listening to the Cisco Optics Podcast, which is part of the Cisco Podcast Network, where you can find other great Cisco Podcasts too. If you want to listen to us on a website, go to optics.podcastpage.io. We also have educational videos on YouTube. Just go to youtube.com and search on Cisco Optics. Thank you for listening.
[00:14:07] This is Pat Chow, Director of Technical Marketing Engineering at Cisco Optics. The next episode is part five of my conversation with John Calvin. See you there.