In The Arena by TechArena - The Future of the Network and AI with MIT’s Muriel Medard

Episode Date: March 26, 2024

TechArena host Allyson Klein chats with Muriel Medard from MIT about the trends in network innovation, how AI is infusing into telco, and the shaping of 6G.

Welcome to the Tech Arena, featuring authentic discussions between tech's leading innovators and our host, Alison Klein. Now, let's step into the arena. Welcome to the Tech Arena. My name is Alison Klein and we're coming from Mobile World Congress in Barcelona and I am so excited to have Muriel Medard back with us on the Tech Arena. Welcome Muriel. Thank you for having me again Alison and it's great to be in person. Exactly, last year's episode was recorded online, so this is a great treat to be able to sit across the table from you.

Why don't you just introduce yourself to the audience and remind folks where you focus from a standpoint of your academic pursuits as well as your private sector pursuits. Thank you. My position is an MEC Professor of Software Science and Engineering with the School of Engineering at MIT. I'm a professor in Electrical Engineering and Computer Science, and I lead the Network Coding and Reliable Communications group inside of the Research Lab for Electronics. TechTransfer, as we mentioned, for network coding, in particular with a company named Steinwolf, which I just came from a demo, and Fred had proof on that. And then Codog, which is a company for transfer of intellectual property, in particular.

So that's what I do. That's awesome. I follow you on LinkedIn. And for the last year, I've just been amazed at all of the places that you've gone and spoke. You are so engaged in this industry and doing such interesting work. Why don't we just start with last year, you talked about the solution that your company delivered in market. And I know that you've made some progress. You mentioned with Red Hat. Can you just remind folks what that solution is focused on?

What problem are you solving? And how is the market response going? Yes, thank you. So a lot of the issues that we've heard about when we've talked about 5G, including the last conversation we've had, are around latency. And you mentioned that one of the issues with 5G

is they've tried to go at the latency problem by just adding more bandwidth, and that doesn't quite do it. And it's also extremely expensive, particularly for the operators, for the ones having to buy that bandwidth. So what we do is we use coding, as the name of the mic would indicate,

and we infect and repair packets ahead of time. So not just after a problem arises, but proactively before the problem arises to fix the problem. And that actually lowers latency massively, really increases the reliability. Sure. And that's what latency massively, really increases the reliability. Sure. And that's what we've been doing.

So, as I mentioned, we have the brand hat. Any people are OpenShift fans, go and check it out. But also, if you go to the booth at Cradle Point, you will see that in their SD-WAN solution, they're incorporating network coding right now also to us. So there's great progress since last year. Yeah, it doesn't surprise me at all that you've seen that progress, especially as you described the value proposition of the technology last year, but I'm delighted to see it.

Now, I wanted to talk to you about 5G. As I approached MWC this year, I was thinking about, you know, is this going to be the year that we see things like broad proliferation of Open RAN or breakthroughs in private 5G? And, you know, you talked so eloquently last year about some of the challenges that the industry has seen with the 5G standard and adoption of certain use cases. What are you hearing from operators about these types of use cases? Maybe also I throw in cloud-native adoption in terms of broad service delivery. What do you see? And do you think this is the year that we're going to see 5G fully proliferate across networks? Yes, I think this is of course one of the core questions

that everybody's asking. You know, I'm not a market analyst, so I may not be the best at specific timing because of course many of those decisions are not only technological they're also commercial and they deal with a lot of different contracts maybe let me just reprise one of the themes that you brought up which is the theme of native i and of course one of the things that's quite striking at mwc this year is a lot of emphasis on ai i think that he has great promise but one has to recall that, you know, let's call it machine learning maybe rather than AI, is about optimization. Some years I actually co-teach a machine learning class at MIT, you know, hundreds of students in your class.

And the way we teach it is always starting from optimization. So it's a means to optimize. Now, we need to have something to optimize. So if you have a network that's going to be monolithic and not have a lot of different possibilities for you to operate too. You may use machine learning to develop very in-depth knowledge of circumstances, even predictive capabilities. But then what are you going to do about it? Right. Certainly, you know, better optimization and having data sets in which you can learn is helpful,

but you're going to be stymied long-term if you don't have the flexibility to then change things and change things in meaningful ways. Right. So I think that the optimization aspects miss, if you will, one half of it. I know some people are dismissive of AI and ML. I've never done that. But I do emphasize that then you need to have the ability to optimize. And if you have something very monolithic and the standards are still very monolithic,

then what are you going to optimize over? If you have one or two choices, then it's overkill to have really sophisticated optimization techniques and then your means to perform that optimization are limited. It's interesting. I wanted you to talk to me about AI and your perspective on it, so I'm glad you brought it up. I think what you're arguing is that taking advantage

of some of the codification of the network is required in order to take advantage of technologies like machine learning. Am I reading you right? Yes, I think that tool to implement it, but again, the underlying technology and the hardware has to also be sufficiently flexible. Right. Otherwise, you're again going to get stuck. So what delivers that flexibility in your mind? You mentioned Open RAN, right? So Open RAN has been, if you will, at somewhat higher layers of the radio stack.

And the idea there is that you have interoperability and, in effect, what I like to call the API-ization. So, you know, if you look at other parts of networking, you have, in effect, standard interfaces. And then because we have standard interfaces, you are able to do a whole lot of optimization within modules. Sure. Because I can do really clever stuff as long as I'm able to interface correctly with the other parts of the system. Right now, if you go below the RAM in 5G, it's still very monolithic. There are obviously options. I don't mean to be dismissive or glib about it,

but it's still, by and large, fairly monolithic. I'm hoping, particularly looking at 6G, that if there's more API-ization, then you can do really clever optimization. And now when you have ML, you have something to optimize it for. Right. Because you can say, well, it's not just that I have to use

exactly this kind of, say, error-correcting code, hybrid ARQ, whatever it is, which is not at all optimized, I have to. So that's really what I'm hoping will lead to the true use of ML. Again, I need something to optimize over. And if you've fixed everything that I do, you've taken my choices away. So having an API-ization so that there are choices, so that you can have real innovation inside these different volumes.

Now, you brought up 6G, so I've got to ask you, what is the state of the standards efforts in 6G? When do you expect to start seeing publications of standards in that space and then how anxious is the industry in adoption of 6g so these are all tricky questions and i'm going to start with my you know disclaimer of the fact that many of these decisions are not necessarily technical. And so I circumscribe my remarks on the more technical side. Certainly there are efforts right now in ATIS, which is sort of a North American 3GPP body, and with what's called Next Generation,

maybe rather than 60, although often people just know it by 60. We have a research council, and I'm a member of that. So there is activity from industry. There's more of an effort, I would say, this time around to try to bring in academia, hence my involvement, as well as a lot of other colleagues. You know, there are publications from NGA, and they're available online.

So I think I'm optimistic, but time will tell. Okay. I think time will tell. With respect to how much pull there is, that's a good question, you know. And at the same time, as you know, there's a lot of, let's say, unfinished work in question. Right. It's just been pumping out. So when does one end?

When does the other begin? These are all very, very fluid situations. You know, I think that one of the questions that I have is, you know, I don't think that anybody really saw artificial intelligence becoming such a course in the industry when the 5G specs that didn't really intend for their use? Or is AI a forcing function for a move to the next standard with 6G? And I'm wondering about your perspective on that. That's a really, really good point. Again, I think that depends on whether 6G does allow the kind of flexibility that then makes the machine learning worthwhile. And to be clear also,

not everything requires sophisticated optimization. So use machine learning for things that you cannot optimize easily. If you can optimize it easily, you don't need it. Some of the low-hanging fruit in the optimization have yet to be realized. Right. And I think a big part of my success is whether people have, if you will,

the courage to undertake a really thorough spring cleaning. Because 5G has so many layers still. And of course, there's some fantastic innovation. because 5G has so many layers still. And of course, there's some fantastic innovation, don't get me wrong, some really clever people have contributed to that. But there are still some layers, which I call them often sedimentary layers,

where you go, oh, I remember you from 3G. What are you doing there? And nobody had, if you will, the gumption to remove it. A lot of it because I think the standardization process is fairly fragmented. And so, you know, you don't want to tell the other committee, why don't you just disband yourselves? We're just going to go away with you. It's a difficult conversation to have. So I think that that fragmentation is starting to become not productive, I would say.

Yeah, not productive. So I think you'll... That's one of the reasons why API-ization is great, right? Because we said, look, here's our interface, and then we're really going to discuss about that interface, and then whatever you do within your module, go for it. Right. Now, it makes innovation a lot more straightforward in my mind. Right. That's fantastic.

Now, all of this stuff runs on silicon. Correct. And I know that you have perspectives on the silicon arena, the tech arena full of silicon all the time, something that we're obsessed with. So what is the latest in networked silicon and how do things like the CHIPS Act influence what's capable of providers in terms of what's being delivered today? Yeah, so the chipset potentially will have, I think, a very considerable effect. It's early still, so we can't, you know, we can't predict these things, as you

know, are, you know, epically tricky right now. But certainly the emphasis towards, you emphasis towards reconsidering the optimization of the hardware, I think can have a large impact. That being said, you still need, you know, Silicon is still implementing algorithms. It's still implementing protocols. It's still implementing things that were designed. And that implementation has to be done properly by the standards body. One of the things I'm hoping will occur, and I'm talking here partially really more as a,

you know, with my professor hat on, is so that they're reconnecting, particularly in the network arena, to the hardware. You know, I'm an electrical engineer, for the most part. And you know, even though I'm a professor of science, I'm actually an electrical engineer. And when we do optimization for instance of algorithms we always think how does this look like when you actually implement a new chip and certainly some of the work we've been doing recently has been implemented into chips we've been doing here about chip work so how you know re-establishing that connection which i think has become somewhat frayed. Just to give you an example, you know, when people talk about complexity,

and you say, well, you know, it really matters whether you can paralyze things or not. You know, we kind of knew it, but people weren't really thinking of it. It's not just, you know, the number of operations. It's also, is this easy to multi-thread? Is this going to get faster or not? Right. You know, are you having to do a lot of reads from memory or not? Those sorts of basic tenets of proper design, I'm hoping to have that reestablished.

You know, it's interesting. As we look at the Silicon Arena, we think about the slowing of Moore's Law. AI is certainly being a disruptive force in terms of the requirements for performance coming from microprocessors. We've seen network infrastructure embrace some industry standard technologies with the advent of NFV and SDN. But some network services are some of the most demanding performance areas in terms of anything that runs on silicon today.

What do you think the future looks like in terms of architectures, in terms of silicon players in this space? Who do you think has positioned themselves well in terms of what they're offering to serve the requirements for the network of the future? That's also a very tricky question. I'm going to give you a weasel answer, but I'm hoping to use the weasel answer rather than specific company. I think that right now silicon companies, particularly on the network side, haven't

generally developed, I'd say, a thin internal capability for networking communications. They have people, and some of them have plain old just divested information. And I think the companies that do have more networking communications have somewhat

divested or never been as much on the silicon side. So I think they'll have to collaborate. I think the ones that collaborate on both sides, those will I think they'll have to collaborate. I think the ones that collaborate on both sides, those will be the ones that do very well. Again, as I mentioned, if you're a chip company

and you're just implementing straight up vanilla algorithms and maybe doing it appropriately, well, okay, good for you, but that's not going to give you an edge. Similarly, if you're doing sort of standards a little bit in a vacuum and not really looking at

what's coming up on the chip side, I'll give you a very simple example from my own research just to illustrate what I'm talking about. We have an error decoding algorithm called Grand Technology Coding that I co-invented with Ken Duffings now at Northeastern

University. We published the algorithm and several groups did synthesis of it. And then we did a chip which we presented last year at ISSCC with rather a digital view. And the reason our chip was so much better than even the synthesis,

usually synthesis is much better because synthesis is a bit in theory, you know, and theory always works better than practice, was because there was really tight work between the people who had, and, you know, when I say better, I mean more than 10 times better. That's incredible. Yeah. You know, we beat the best in class chips right now by

over a factor of 10. But even the synthesis, we were about four times better. The reason was they did an excellent job just taking our algorithms and kind of implementing them for all, right? We were working with

the actual circuit speed and going, just to give you an example, oh, you know, when we say sort, it doesn't mean you have to sort everything. You can sort as you go, right? Because we have a more subtle understanding. Some of the things we did, for instance, we had a memory computation, right?

And that sort of discussion actually gives you a huge amount of gain. Right. And right now we're missing it. You know, you talked about Moore's Law. If you look at the original Moore's Law, it wasn't just about improving silicon. It was also, one of the aspects that Moore talked about was algorithmic development. Right. talked about was algorithmic developments. So he understood that it was a combination and a synergistic combination

of ACW circuits, of course materials, but also algorithms. And right now the algorithms are sort of floated off a little bit of, you know, to the side by themselves. So bringing the algorithms

back to the circuits, I think that that's really side by themselves. So bringing the algorithms back to the circuits, I think that that's really, really powerful. So that's an example. So I think that the companies that are able to create those synergies between the people doing, I'm calling it algorithms, but a lot of the better work, and that's the technology on, if you will, the descriptive side, the math side, and the people who are doing the circuits, that's where those are the ones that are going to be successful.

And right now, I don't see a company having both in-house. Right, right. I just don't. You know, as you were talking, one thing that came to mind is this is the reason why chiplets exist. Correct. And that a future of really tight chiplet design where you're getting best-in-breed type of capabilities and utilizing open architecture, whether it's UCIE or, you know, future form factor standards that are being discussed in the chiplet space, really would bring incredible innovation to this space.

Absolutely. And actually, you know, our plan for the grand encoder right now is to put it into, is actually to make chiplets, to put it into the prolonged codes. That's exactly what we're doing. And chiplets are a beautiful example of what I was talking before about the APIization. Right? It's like, you know,

your chip here. Here are the inputs, here are the outputs. This pin better do this, and this pin better do that. And then what you do inside, go for it. That sounds exciting. So, Muriel, how do we actually get this magical land of network chiplets to come together?

Do you think the industry is headed in that direction or do they need some nudging? Both. I think certain people recognize this, but, you know, there's a lot of, I would call it understandably, inertia, right? I mean, this is an industry that has a long history, and some of this approach may be disruptive to existing business lines. You know, when do you actually decide to let go of a particular approach and embrace the new approach? Those can be very, very tricky

transitions. Now going back to the CHIPS Act and policy, obviously semiconductors have never seen so much focus from broader industry and broader even the political sphere. Does that

help put pressure on the industry to move to these new models? I think it would be the better, you know, the better recipient of that question. My non-expert belief or let's say my observer belief is yes, but it's not clear, right? Because there's a lot of temptation to just do what you keep doing. Particularly if you don't have in-house the expertise to tell you otherwise. Right. As I said before, I don't see right now companies necessarily that have, I mean, they'll have a few people here and there, of course,

but that fundamentally have in-house code expertise. Yes, that makes sense. I'm going to shift gears. We're in the middle of the FIRA in Barcelona. Yes, we are. People can't see us, but we have thousands of people milling around us as we're sitting here. You're engaged in conversations with people from across academia and the industry here. What are your key takeaways thus far in day two of MWC? And what are you most excited about, about what you've seen?

Yeah. You know, it's funny. I haven't quite distilled them myself, you know, partially because it is such an overload that we were talking about before starting the podcast. Certainly the softwareization part that were talking about before starting the podcast. Certainly the centralization part that you talked about before is very tantrum.

And I think certainly I'm seeing more openness into discussing types of collaborations that maybe, I would say, when three years ago would have been unusual or fringe or just downright, you know, weird. Weird.

Yeah. Unexplainable. So I'm seeing a little bit more of that idea of having those conversations. And, you know, coming also as somebody who participates very actively always, you know, academic type on. I am seeing a little bit more openness, I think, to ideas from academia. So I think that the idea that, you know, everything's just

going to be done in-house and that's it. I'm seeing a little bit more of a let's say a little less reluctance to engage active conversations. That's wonderful. Well, thank you so much for spending time out of your very busy schedule. It's been a pleasure.

Same here. Thanks for joining the Tech Arena. Subscribe and engage at our website, thetecharena.net. All content is copyright by The Tech Arena.