Transcript | Quantum Interconnect with Entangled Networks

Things get tricky when you try to cram too many qubits onto one quantum processor. Is the best way to build a machine with thousands of qubits actually to connect or interconnect many smaller processors together? And can this concept be extended to interconnect several large-scale machines across vast distances? Join us for a look at the technology that might get us to behemoth-sized quantum computing in this episode of The Post-Quantum World.

I’m your host, Konstantinos Karagiannis. I lead Quantum Computing Services at Protiviti, where we’re helping companies prepare for the benefits and threats of this exploding field. I hope you’ll join each episode as we explore the technology and business impacts of this post-quantum era.

Our guest today is the CEO of Entangled Networks, Aharon Brodutch. Welcome to the show.

Hi, yes — nice to be here.

Yes, I’m super excited to have you on. I mention interconnect so often on this show that it could be the basis for a very dangerous drinking game. So, before we dive into the amazing potential of interconnect, please tell us a little bit about Entangled Networks.

We’re an early-stage startup. We’re based in Toronto, and the founders, me and Ilia, we’re both physicists by training. My background is very long in quantum computing, and then we have a third cofounder who has done this multiple times before, including building a huge optical telecom company. We decided to solve the biggest problem in quantum computing, and that problem is trying to scale up, trying to build a huge quantum computer. The only way to do that is to take multiple small quantum computers and connect them together, and we’re building everything that’s needed to do that.

That’s amazing. That’s part of the reason why I keep bringing this up. So, yes, let’s take a step back, and we’ll just examine some of the reasons why you want to work on interconnect.

More than just chaining together those multiple big computers out there, it can help us increase, as you hinted, qubit counts inside individual quantum computers. I believe there’s already a bottleneck in having too many qubits in one quantum processing unit. Do you want to talk about that?

Yes. We do have a huge bottleneck, and I always like to use IBM as an example because they have a ton of information that’s publicly available. If we look at what IBM has and what they advertise, they’re big — they have these two big ads. One is, “We have the Eagle processor.” It’s 127 quantum computers, and then — 127 qubits, sorry — I wish it was 127 quantum computers.

Yes, I know. Shorter queue times.

Yes. They have this other big advertisement, and very recently, we saw this on Twitter: the 512-quantum volume, which is, yes, quantum volume of 512, which I would call quantum volume of 9. They do like to take 2 to the power of the actual qubit count. And that’s done in a 27-qubit machine, so that’s done in Falcon. So, you’re seeing that they have this huge machine — it’s 127 qubits, it’s very cool — but actually, the best-performing one is a 27-qubit machine, which is, from a qubit count, two generations further back.

So, adding more qubits to the machine is just insanely hard, and it’s insanely hard for two reasons: One is just to add more qubits to the machine and get the controls done, and that’s what the 127-qubit machine shows. It shows that they can get the controls in to get 127 qubits. That’s amazing, and it’s true: Every single technology will have this kind of problem to try to control that many qubits.

But then you want the qubits to interact. That’s fundamental for the quantum computer. But you want them to only interact when you’re telling them to interact, and they don’t only do that. They start to interact when you don’t want them to do it. It’s like packing a ton of kids into a classroom: At some point, noise levels just go up, and the teacher stops being effective. So, we’re trying to do both of these things at the same time — it’s insane. At some point, you have to say, “This is an optimal size. Let’s start connecting modules together.”

Interesting. So, you’d have the kids in a bunch of little pods, and then the teacher moving between them. Let’s put it that way.

In a school, you want different teachers sitting in different classrooms. You don’t want all the kids in one giant classroom with a teacher teaching each group, because they’ll start talking between groups. And it’s exactly the same.

Yes, that’s a great example.

PsiQuantum, when they talked to us, they talked about this idea of building one large computer from interconnected modules, but they weren’t ready to give any specifics on how many qubits would be in each of those modules. I tried — I was thinking, would it be 50 in each one? I wasn’t sure — but their goal is a million total. You’ve got to think there’s going to be a good number of modules in those things. Do you have any thought on what the ideal number of qubits might be per module?

Yes. That’s a super tough question, and the reason it’s so hard is that it depends on a ton of different factors. You’re looking at the actual qubits and how they perform, you need to look at the interconnect and how it performs, and you need to think of overheads, because the interconnect does take up some of the qubits. You have to take all of these things together and then also look at the kinds of algorithms you want to run at the end. If you’re looking short-term, it’s NISQ algorithms and small error-mitigation techniques. And if you’re looking large-scale, it’s the quantum error-correction code that’s running, and you have to optimize everything for that.

And then, when you look at PsiQuantum specifically, they have had this incredible idea, which is completely unique in quantum today, which is, they were thinking about scale from the get-go and about manufacturing. We don’t think that much about manufacturing if you’re looking at most of the companies. When you think about manufacturing, you’re adding this other issue to the equation, which is, what’s the best size to manufacture? There is a tradeoff, which is just about costs — the number of defects you have before you have to throw something out. They added that Day 1 to the equation, and so now you have to think modular. You cannot — you do not want to — manufacture a device that’s in qubits. It’s definitely very hard to guess what the numbers are, but, yes, a ton of different factors go into making those decisions.

Would you say it’s the type of qubit that represents the limitation of each module? Like PsiQuantum’s photonic, IBM’s transmon — but IBM publicly already said that their pathway to 4,000 qubits will involve some kind of interconnect. And they’re going to roll it out in three stages, how they’re planning on doing that, but their goal is ultimately to admit that, yes, they’ve got to have multiple processors working together to create that many qubits. Do you think it’s the type, also, that limits?

The type of qubits will be a huge factor, and not just, is it photons, is it superconducting, is it ions? It’s the specifics of the system. It’s the specifics of the interconnect and how efficient it is.

If we want to have an idea of numbers, people have been tossing some numbers around, not for photonic but for ion traps. The ideal number seems to be, don’t go over 50 ions in a single chain — it just adds too many difficulties, and so you want less than 50. Ideal seems to be around 20 to 30 ions in a single device. But what you can do with ions is, you can have multiple chains on a single device, and you can shuffle ions from one to the other, and a lot of companies are thinking about doing that. That’s a good first stage.

But once you take ions and you literally move them from one side to the other, you can imagine that you don’t want to move them too far. Now, you have this big connectivity-constrained problem. You’re only talking to chains that are very close to you. At some point, that just doesn’t scale very well. You want to start getting outside of that system optically, and again, numbers are being tossed around — we’ll get to about 1,000 to 2,000 qubits. Let’s say somewhere between 500 and 4,000 qubits would be the number before you have to go out optically with ions. I would guess that the lower end is probably the more efficient, but the higher end is probably where people want to start doing it just because of other engineering constraints.

At around that point, you think, trapped-ion, the only way to go up will be to chain entire computers together.

With ions, that number with superconducting, like you said, we have numbers from IBM. They do say, “We push up a single module to a thousand,” and then they’re looking at multiple ways of doing even that. They’re thinking of breaking down that module into three small modules that are still on the same chip. And then, at some point, again, you have to go into an interconnect.

Yes — Neutral Atom is kicking around going for a thousand qubits next year, too, with their Pasqal, for example.

Yes, they’re super optimistic. The neutral ions, they do expect to get to about 10,000 before going optically. And here is this tradeoff that you need to think about: It’s very possible that they’ll reach 10,000 — an array of 100 by 100 — but now you’ve got this grid with interactions that are at a fairly short distance, and you have to start optimizing. Is 100 by 100 the optimal size for a processor even if it’s the biggest one you can aim for? Very likely, that that is not the optimal number. They might initially start trying to reach 10,000, and then at some point realize, “From an architecture point of view, we should have smaller modules.” So, still unknown — we haven’t seen a ton of developments with neutral ions, although it’s an early technology and it’s super exciting.

Let’s dig in to the difference between interconnect when it’s inside and outside. Everyone listening to this has multicore CPUs in their lives. They have laptops, phones — those are usually multicore in some way. Can you explain some of the ways that quantum interconnect inside is different than classical connecting of these CPUs together?

With the classical CPU, you need to transmit information from one CPU to the next. There’s a bottleneck there. It’s slower than the CPU just working on its own, but it’s fairly simple. We know very well how to transmit the information from one side to the other.

With quantum, you have to preserve coherence. I’ll say one word about what that means. The number of possible states in the quantum computer is exponential in the number of qubits, and what you want in order to get the power of the quantum computer is superpositions of these states, and maintaining these superpositions. Let’s call that coherence, for simplicity. What you want is to keep maintaining these superpositions as you transmit information.

The problem with that is that anything outside the quantum system itself destroys coherence, and so one big thing is measurement, and so you can’t measure. But the other big thing is, essentially, any interaction with anything outside is going to reduce coherence. So, sending information outside your computer and hoping it will reach the other side safely is a problem. You can’t copy, so you can’t try and keep on repeating until success. You’re in big trouble, and that’s where entanglement comes to help us.

So, you can create an entangled state. That entangled state is a quantum system that is comprised of two subsystems, but they act together, and so what you do on one side influences the other side — what Einstein called a spooky action at a distance. It’s very hard, it’s very dangerous to start playing around with this idea because — let’s call it a mathematical idea that the physical implications are a bit hard to imagine. But one thing that it allows you to do is transmit quantum information across this gap using an entangled state and without actually moving the quantum information physically around from one point to the next. And one way to do this is quantum teleportation. Again, that uses the entangled state.

What a quantum interconnect does very differently from classical interconnect is, it doesn’t move information around. Instead, it generates entangled pairs. So, what you want is to generate entanglement between these quantum cores, and then you’ll use it later to transmit information.

Entanglement is a core part of doing quantum computing. It’s one of the things you need. You have to have these things to be able to be entangled. When you’re entangling from one processor to another, you are, in some ways, making them behave as one processor.

That’s right. The quantum state can be superposition of the entire space that includes all the qubits on one processor and the next one.

That’s a huge difference, because when you do interconnect with classical CPUs, you’re just set. You’re almost sending answers instead of the working equation. You’d be sending that to the next machine. Whereas in this case, you can have the two parts of the two CPUs — QPUs, I should say — work together.

So, that’s how it is when it’s close up. Now, if you wanted to do this farther apart, does anything change? You’ve got discrete machines. Let’s say PsiQuantum does build a million-qubit machine with all the modules inside. How would it be different to then connect it to another million-qubit PsiQuantum machine? What would that machine-to-machine interconnect be like? Would it be very different than the closeup one?

It’s a very different use case from the user’s experience. When you think of a single computer, from your perspective, when you’re using your laptop, which is four cores, eight cores, you don’t want to keep on thinking, “I have an eight-core machine.” You want to think, “I have a computer — it should work.” So, from the user perspective, I would look at this million-qubit device, which is built up of 10,000 100-qubit subsystems.

In the case of PsiQuantum, it gets much more complicated, but let’s keep it simple here. You don’t want to start thinking about that, and so that entire machine has to work seamlessly as one machine. That is a very different problem from saying I have two computers and they’re communicating like I’m communicating with you right now. You still need that entanglement in order to facilitate communication for the same reasons that I explained before, but now you need to think less about rates. You might be less worried about loss because you don’t necessarily need them to be working together. Although, in many cases, you still want the entanglement to be there, but it’s very different talking to each other than working together as a single machine. And we know this classically. It’s very different thinking about an internet and thinking about multicore computing.

Yes, it’s a different kind of bus there. Are there distance considerations when connecting systems like this? Would you be limited by the length of something like a fiber?

Yes, so, definitely, distance, again, we can look at the classical world and see distance plays a huge role. You do get losses as you go longer. When you think about connecting a single computer, distances are relatively short, and so you deal with losses as you get a little bit further apart, but it’s not end of the world. When you think about internet-type communication, losses are going to be a huge problem. So, now you have to put on repeaters, which replace amplifiers, which classically allow you to amplify the signal. Repeaters allow you to do long distance. And thinking about a small network, we don’t need repeaters, so distance is less of a problem.

Let’s say if it was close together, would it be possible to link computers of completely different types together — let’s say a neutral-atom and a transmon? Would there be an easy way to do that with the solution you’re working on?

Yes. If I would guess what’s going to happen in 20 years, I would say quantum computers would not be a single modality. There are advantages and disadvantages to each modality, and we will see computers of completely different types working together. But that’s long in the future, and the first-order reason why it’s hard is, you need to convert between wavelengths.

Superconducting qubits, they work in microwave. Ions, they work in invisible. Even converting between different ions, between the wavelengths of different ions, is hard. It’s a much bigger challenge to convert from microwave to optical.

On the other hand, that challenge has to be met, because if you want to connect multiple superconducting or anything that works in microwave — the kind of approach that Intel has in solid state — anything that works with microwave, you need to convert microwave to optical if you want to get outside the fridge. This is an absolutely necessary technology for IBM’s road map. It’s still science. It’s not engineering quite yet, but that science has to be solved very quickly.

We’ll have the building blocks. Once we have that, we already know how to convert inside optical. Once we have that, yes, it’s going to be just engineering. It’s huge, but we’ll make huge strides.

The first system-to-system interconnect we’re going to see is most likely going to be of the same type of system. It would be two photonic machines, or whatever, talking to each other, not a mixed bag yet.

Yes, a hundred percent. If you think about photonic — if you think about anything that’s not atoms — even that is very hard, because the photons, even if they’re the same wavelength, they’re not exactly the same. Ions and atoms tend to be different because they’re all exactly the same.

You would think that the most successful interconnect first system-to-system would probably be trapped-ion or neutral-atom, or something like that first?

Yes, or with diamonds. If they mature as a computing technology, yes — things with atoms inside, yes.

It’s great to get that inside look to how you see the progress mapping out. And your company is going to be publishing a blog — it will probably be up by the time this hits. You were kind enough to share in advance. I’d love to talk about some of the results you share in there, and name this version of interconnect that your team’s working on, and it’s called Qlink.

Yes, our first-generation interconnect, we call it Qlink A. We plan on having a B, C, D, etc., version, and that first version will connect two quantum computers together. Again, thinking in terms of classical networking, it’s a network adapter. It allows you to extract that information as photons and use that to establish entanglement.

And then, we want to start educating the world about how we’re going to see interconnects in the future, how information gets transmitted across. We’ll have a series of blog posts slowly explaining how these things are going to work. From the very specific protocols, teleportation is one example protocol that can work. There are things that are cleverer and nicer than that.

And then we’ll show some results on our plans for hardware, but also on our software, which exists and is running. One set of results shows where the advantage is with the interconnect, even if that interconnect is not fantastic. That’s the big problem currently: Interconnects are very slow stuff — that has been demonstrated in labs.

The device we’re building, it’s going to be a bit faster than what’s been demonstrated, but it’s not going to be orders of magnitude faster, and so you still have to deal with that bottleneck. We have results showing that even in that case, thinking of today’s quantum computers, you’re still getting an advantage, and it’s not trivial.

And then, we’re also showcasing our compiler. So, what does a compiler for a multicore system do? It reduces the number of times we want to use the interconnect, because the interconnect is a bottleneck. You don’t want to cross that bridge too many times. And we have software that does that extremely well. Again, bringing the viability of the interconnect much closer in terms of what the performance metrics we need are.

I’d love to hear the good and bad sides. First, we’ll start with the overhead hit that you’re seeing in these initial experiments. I don’t know what metric you’d want to use, but what kind of hit are you taking when you do cross that bridge, as you mentioned?

At the moment, what’s been demonstrated in labs around the world is interconnect speeds of about 200 Hz. That is very slow. We can reach about 1 kHz — or we should be able to reach close to 1 kHz — on ions. Again, everything is very specific to the underlying system.

Ions, they work. A standard 2-qubit gate is about 100 microseconds, and so you’re getting about an order-of-magnitude hit in terms of the time it takes the interconnect to work. It’s not great, but it’s workable. What we’ve found is that that you can still operate with that.

Have you already been able to show that with two systems of a certain size, you can see some benefit even with that order-of-magnitude hit in the transmission?

Yes. In simulations, we can show that if you take two small quantum computers and you use an interconnect, you will get an advantage. One example that we’re going to have out soon is taking two 6-qubit quantum computers and trying to solve a 12-qubit problem on it. There are tricks where you can try to use those without an interconnect, and you get a certain result. And then you say, “What happens if I use an interconnect?” In that case, everything is perfect. I have a 12-qubit machine. But what if the interconnect is not perfect? Now, you’re taking a hit each time you’re using the interconnect. And we model that, and we still see a huge advantage from using the interconnect. If I was to take our design and plug it into a machine today, we would see a performance advantage from using the interconnect.

That’s terrific. If you were to stretch this out a bit on the B, C, D versions of Qlink, you can imagine that if you took two 27-qubit processors, you would see performance that would rival the 127.

The big catch is, if I was to take two 27-qubit machines, these two IBM machines, I could probably squeeze a little more power from those machines because they’re interconnected. But those, even if I used magic, I would not get the entire 54 worth of qubits. I wouldn’t double the computational power, because the processor itself is not there yet.

What we need is, ideally, connecting a large number of even smaller processors where each processor is really good. That’s where industry will be aiming very soon — getting a ton of computational power from smaller processors. There are, again, a lot of engineering considerations that don’t come out in just seeing performance, and this is one of the reasons you’re seeing bigger systems: Not because you try to get more performance — you’re trying to test out these new engineering ideas.

We have to start somewhere. And this approach is a beginning. Is there an ideal goal you’re shooting for? Ultimately, how good do you want Qlink to be? What kind of overhead losses? Do you envision it being — like, what speed, or whatever? It’s hard to extrapolate, I’m sure, with this, but I’m curious about what kind of ideal you have in mind.

The ideal is, you want to be working at roughly the same speed as a 2-qubit gate. It’s never going to be quite as good, but you don’t want to be much worse. And yes, at that point, it’s a bottleneck, but it’s not a huge bottleneck. So, that’s one very ambitious goal that is going to take a while.

We don’t quite have the physics there yet. We know how to do it in theory, but no one has demonstrated a working interconnect at those speeds with the kind of requirements that need to be there. For ion traps, it’s putting the ions in cavities, for example. And in the shorter term, we want to see the best-performing quantum computer being a multicore quantum computer, and within five years, we’ll be very close to start to show that. Whatever highest quantum volume or whatever benchmark you’re going to take, that will be done on a multicore machine.

It’s exciting because within five years, we expect to see some pretty impressive machines, so pairing up a couple of them might be pretty impressive itself.

You would think, then, that Qlink will be commercially out there within five years. Is that the idea?

Commercially out there, we’ll probably be before 2025, 2026, if industry timelines are what they’re advertised to be — 2025, 2026 is when industry needs to start linking computers together. Now, from when you start doing it — from when we sell our first product until you’ve actually built the computer — that takes a bit of time. So, in five years, we’ll be seeing that computer operate at the highest levels, and this is for our technology, which is for everything that’s not photonic.

If PsiQuantum delivers and they have an incredible team, we’ll see a multicore quantum computer within a couple of years being the best quantum computer out there — fingers crossed that they can deliver on promises. Their machine is inherently a multiprocessor machine.

Yes, everyone’s excited to see that one, and I’m excited to see what you guys come up with. You have 25 more letters in the alphabet to go through for your next versions, so maybe you’ll get lucky and you’ll strike gold on the Q version, because that would just be funny, you know? Because everyone loves to use that letter.

That’s a great overview of this. Is there anything else you’d like to share before we wrap up?

Quantum computing is a strange technology in that we’re doing a lot of physics. As the industry is growing, we’re starting to do a ton of engineering, so we’re doing both of these things in parallel. Getting physicists to think about engineering is not super easy, so we’re learning as we go along. And what we’ll see is a bunch of engineering challenges being done in parallel with physics for interconnects. That means that we’ll see the engineering challenge of getting a multicore quantum computer to work and deliver results on the one hand, but then, on the physics side, we’ll get those interconnects to work at the speed that we would like to see where there is not this insane bottleneck on the system.

Thanks so much. It was great having an inside look at how this is actually developing and evolving right now, rather than just hearing from the outside. I’m glad I got to meet you at IQT, and thanks so much for coming on.

Yes, thank you very much for having me on.

Now, it’s time for Coherence, the quantum executive summary, where I take a moment to highlight some of the business impacts we discussed today in case things got too nerdy at times. Let’s recap.

Entangled Networks is a startup in Toronto focused on solving the biggest problem in quantum computing: scaling. The company feels that the best way to make a high-qubit-count quantum computer is to connect, or interconnect, multiple small quantum processors together. The technology can also be used to interconnect those large quantum computers across many miles and yield even greater scale. Imagine connecting five 1,000-qubit machines, for instance, in different cities and having them act as one 5,000-qubit system.

There is currently a bottleneck when trying to put too many qubits on one chip or processor. Control electronics become difficult to pack, and qubits may experience greater noise levels. Entangled Networks says the best approach is to find the optimal qubit number, build a module with that many qubits and then interconnect them.

The optimal number is still up for experiment and debate. Estimates range from 20 to 50 on the lower end to as high as 1,000 to 10,000 qubits on the high end. PsiQuantum is building its computer with an interconnected model, and even IBM admitted some form of interconnect will be needed for its future systems.

Classical systems already work with multiple connected cores. Quantum interconnect makes this concept trickier because you have to maintain coherence. Entanglement is used to transmit information by generating entangled pairs between modules and then having the superposition of all the qubits act as one system. The gap between these close-by modules doesn’t matter if this is done correctly. Things do get tricky if this is done over quantum networking, with its challenges of repeaters, noise and losses.

In the future, we may be able to interconnect heterogeneous types of quantum computers like optical and superconducting. But in the near term, Aharon feels this will be difficult. Even sticking to chaining up only the same type of processors, though, the benefits can be huge.

Entangled Networks is using something called Qlink to accomplish interconnect and published exciting results that already show ways to gain some advantage today. Qlink interconnect speeds are at about 1 kHz, which is an order of magnitude slower, or a performance hit, on normal 2-qubit gate speeds. But even in this early stage, a demo experiment shows that two optimal 6-qubit machines with interconnect can rival a single 12-qubit machine, so they’re on the right track. The goal, or ideal, is to get interconnect to work at a speed that’s close to a 2-qubit gate. Aharon feels that interconnect will be in use within three years and that within five years, the best-performing machines will be built this way.

That does it for this episode. Thanks to Aharon Brodutch for joining to discuss Entangled Networks and their work on interconnect, and thank you for listening.

If you enjoyed the show, please subscribe to Protiviti’s The Post-Quantum World, and leave a review to help others find us. Be sure to follow me on Twitter and Instagram at @KonstantHacker. You’ll find links there to what we’re doing in Quantum Computing Services at Protiviti. You can also DM me questions or suggestions for what you’d like to hear on the show. For more information on our quantum services, check out Protiviti.com, or follow ProtivitiTech on Twitter and LinkedIn.

Until next time, be kind, and stay quantum curious.