Podcast with Aharon Brodutch and Ilia Khait, Co-founders of Entangled Networks

Aharon Brodutch and Ilia Khait, co-founders of Entangled Networks, a company building an optical interconnect between QPUs are interviewed by Yuval Boger. Aharon, Ilia and Yuval talk about the efforts required to scale up quantum processors, their optimizing compiler, building a distributed quantum computer, and much more.

Transcript

Yuval Boger: Hello, Aharon. Hello, Ilia, and thanks for joining me today.

Aharon Brodutch: Hi.

Ilia Khait: Hi, thanks for having us.

Yuval: So, who are you, and what do you do?

Aharon: Yeah, so we're Entangled Networks, and we are solving the most important problem in quantum computing, and that is the problem of scaling up. At the end of the day, you need a large quantum computer to run things on, and the only way to scale up to really large sizes in a sustainable way is to take lots of small quantum computers, connect them together, and build a large quantum computer made up of many nodes. And we are building the enabling technology for that. And that's a piece of hardware, a quantum interconnect that allows you to connect these quantum computers together into a network, but then also the software that allows everything to function together.

And what we've realized over the last two years is one, people know very little about the software layer. There has been a lot of work done on trying to improve the hardware, which is incredibly difficult, but very little work has been done on the software. And we are able to show that you don't need the hardware to be extremely good as long as you can have good enough software tools to kind of function and use the resources in the best way. And this is where we really shine. And just personally, about us, so I'm Aharon. I'm the CEO of Entangled Networks. My background is in quantum. I've been doing quantum since about 2007.

Ilia: Yeah, so I'm the CTO of Entangle Networks. My background is in condensed matter physics, with expertise in numerical algorithms, many-body systems, strongly correlated particles, things like that.

Yuval: So, let's dive into the first part of your answer about the ability to scale. IBM says that they have a development roadmap. They seem to be sticking by it. They have more and more qubits. Quantinuum is saying that they're increasing their quantum volume tenfold every year, I think. So couldn't we just wait until there are enough qubits? Why do we need interconnects for smaller computers?

Aharon: Yeah, so if you look at IBM's roadmap, then they're going to scale up to about a thousand qubits before they start connecting machines together or devices together into a larger device, and so very clearly on their roadmap. And if you look at IonQ's roadmap, it's the same. At some point, very clearly, you go into optical interconnects that allow you to go outside of your small quantum computer. If you look at Quantinuum, they're taking a slightly different approach in that they're going to use shuttling. So initially, you're still going to have a distributed machine with lots of different small processors, but they're going to use shuttling to move ions from one side to the other. And that can take you to a certain extent. But beyond a few thousand qubits, you are going to want to go optical.

And maybe one thing that you're seeing in terms of how technology is developing is companies are doing two things. One, they're increasing the number of qubits, and two, they're increasing the capabilities of the gates or the capabilities of the per-qubit device. And as an example, IBM has an insane 127-qubit machine, which is an amazing engineering task, but that is not their best-performing machine. Their best-performing machine is a 27-qubit machine, and that's mostly because it's much harder to improve two-qubit gates on a smaller device than on a larger device.

Yuval: So, how does this work? And does it work, the interconnect?

Ilia: So first of all, it works, and experiments in late 2019 show the connection between two, let's call it distant, qubits. So by distant, I mean the qubits that sit on two different devices, and maybe the rate was not super impressive. It was below 200 hertz. But what we show is that, first of all, we can improve that rate, and we show that by using the methods that Aharon described earlier, which means by better compilation and by very clever real-time policies, we can actually make that link work and algorithms perform.

Yuval: And regarding the first part of the question, how does it work?

Ilia: Yeah. So basically, what you need to have, and the most important piece, is entanglement, hence, our name, right? But generally speaking, what you need to have is a channel to exchange information, and that channel is entanglement between two qubits on each... so a single qubit on each device. And that allows you to perform a single operation, a single two-qubit operation, if you wish, on that device. It's a single-use link. So once you use that, you need to produce another pair, another piece of entanglement if you wish.

Yuval: Do you envision this being used primarily to interconnect pieces of the same computer? In your example, Aharon, you said, "Well, IBM may have a thousand-qubit chip one day, but then to scale up they would need to interconnect them." Or do you see this more as a distributed supercomputer-type, meaning I have multiple quantum computers in different cities, and I just want to network them together to get a larger number of effective qubits?

Aharon: Yeah, so I think there are two questions that you asked there. One is about interfacing different modalities, and the other one is about interfacing a large network. There is an added problem if your network is bigger. If it goes across distances in it, you need quantum repeaters in order to make the channel more reliable. So we're not looking at that for now, but certainly, at some point, we would like to have distributed computers. If you're thinking over long distances, what you want is not speed, right? If you have a supercomputer, you don't want that supercomputer to be distributed around the world. You want it at one location. And it's the same for a quantum computer. But what computers distributed over vast distances give you is added security. You can do things like delegated quantum computation and things like that. And you can also exploit resources that you just don't have on hand in your one device.

But this is much further into the future. If you're thinking of a network, first get that network to work on a small scale, right? First, get something to work in a room, and then you can start thinking about the next step. So that's about a quantum internet kind of world. And certainly, that would be very interesting. The other one is about different modalities. Again, there is a huge technological opportunity there, but it is extremely difficult to interface even similar modalities with slight differences in the architecture. For example, using two ion traps with different types of ions, you need to convert wavelengths. And if you're trying to interface to an ion trap with a superconducting qubit, you need to go from microwave, which is where superconducting works, to optical, which is where the ions work. And doing that is incredibly difficult.

And so, for now, the first challenge is getting just one distributed quantum computing system to work. It's a challenge, and it's a need for the industry. It's just hard to build a large quantum computer without this solution. And so everyone has to race very quickly to solve this problem. But once you've solved it, you can start thinking about the next problem, which is how do we get a superconducting qubit, which works extremely fast, to interface with an ion qubit, which has an extremely long lifetime. And so now you can exploit sort of the advantages of these different architectures in a way that's kind of similar to a computer that you have today where different pieces serve different purposes.

Ilia: But unlike in classical computing... In classical computing, you have standards. In quantum, we're quite far from that at the moment.

Yuval: Today, a transpiler has to do a fair amount of work to optimize the circuit to a given hardware, for instance, to minimize the number of swap gates and so on. If you have operations that happen on both sides of the entanglement link, and you mentioned that they might be slower, lower bandwidth, then you probably need a software layer that says, "Okay, how do I parse that circuit so that part of it runs on one computer and one on the other?" How is that happening? Is that part of your offering?

Ilia: Right. So that is the main offering or the main piece of the compiler, that we want to reduce the number of inter-core, inter-QPU, quantum processing unit operations. But in order to facilitate that, what we also do is better distribute the qubits in the first place. Let's say, pieces of the quantum circuit or pieces of the code are distributed in a smart way to begin with such that these small instances can run on different cores. But again, there's also... A very important piece is, what do you do in real-time? But that is what happens after the compiler.

Aharon: Yeah. So maybe on the compiler front, I can add a little bit in it. Initially, what we started off with is we said, "From a theory perspective, we don't need a ton of entanglement in order to get a huge advantage from a multi-core processor." And so we tried to look at algorithms, and we saw that the number of times you'd have to exchange information is drastic. It's huge. And our reasoning was this is probably a compilation problem. It's probably not working because the compilers are not built to solve this problem. And so our first piece of software was a compiler that does this so that we can start internally testing and understanding how many times you need it to transmit information from one side to the next.

And so we built a compiler. And as we're building this compiler, we also try to test it out on some other devices that are not distributed quantum computers, and so the things that exist out there. And we saw that we've actually hit on a nice way to do compilation that is independent of a distributed system, but it relates very much to connectivity constraints that you have on most architectures today. And so, really, we have this incredibly nice tool. And just to kind of understand how important compilation is, we're seeing up to 90% decrease in the number of interconnect operations when we take our compiler versus some other commercial compilers out there that sort of try to brute force, minimizing the number of two-qubit gates. And so once you look at it as a distributed problem, you are also reducing the number of two-qubit gates. We are getting that extra advantage that I was mentioning earlier.

Yuval: You mentioned earlier IBM's 27-qubit and hundred or hundred-so-qubit machine, but they have different gate fidelity or different two-qubit gate noise model. Have you simulated the effect of scaling up? I mean, could you, for instance, say, "Well, if you networked multiple of these 27-qubit chips, you would get something that's better than the hundred-qubit chip?"

Aharon: Yeah. That's a fantastic question. This is something we're doing right now. There are limitations to what you can simulate classically. And so right now, we can do these kinds of estimates on smaller instances where we can do a complete simulation of the architecture. We can say, "What would happen if you took two 10-qubit machines and connected them together with certain specifications that we can get from IBM's website or from Azure or from AWS?" And we see definitely the interconnect doesn't have to be extremely good. That being said, with superconducting architectures specifically, there are no good interconnects at the moment. So it's going to be a much harder challenge for the industry as a whole to connect superconducting machines together. But IBM has promised that they're going to do it, and IBM has so far delivered on their promises. So I am optimistic about us seeing superconducting devices being connected together in the not-too-distant future.

Yuval: Tell me a bit about the company. Where are you located? How large is the company, at what stage? How close are you to the working product?

Ilia: So the company consists of seven physicists. It was founded by me, Aharon and Rafi. Rafi Gidron is a serial entrepreneur. One of his biggest claims to fame was Chromatis Networks which were sold for about... What was it?

Aharon: 4.5 billion.

Ilia: Yeah, a lot of money to Lucent. That was the first Israeli unicorn I guess. Other than that, as I said, there are seven physicists here in Toronto, Canada. What else?

Aharon: Yeah, we've existed for two years. We're funded by OurCrowd. We received the funding in May 2021.

Yuval: How close are you to having a working product?

Ilia: Yeah, that's a great question. So as we pointed out at the beginning of this, there's no USB kind of a plug-and-play system in quantum, especially for interconnects. Therefore, our biggest challenge at the moment is to find the right partner to partner with and to build that system specifically for their devices. Having said that, we have a detailed blueprint for a prototype, and we're beginning to engage with partners to build it right now. So I'd say about a year from now, we'll have our first actual prototype for a demonstration.

Aharon: And that's on the hardware front. But then on the software, like I said earlier, we have a compiler. It works. We're not making it available to the public yet, but it works. And if anyone is interested in trying out multicore algorithms, we're happy to discuss and show off the abilities there.

Ilia: And just to add some more meat on the compiler front, as Aharon pointed out before, we do see a lot of improvement on monolithic devices as well. You can test that without going to a non-existent, at the moment, multicore architecture. So there are definitely advantages to using that today.

Yuval: And is that the business model? Is it to partner with a hardware vendor that makes quantum computers and say "We have a better interconnect, and this is the best way for you, the hardware vendor, to scale up your computer?" Is that correct?

Aharon: Yeah. So our business model is to be the Mellanox of quantum computing doing exactly what they did in classical computers, providing the hardware and the software for making supercomputers. And we're doing the sort of basic network infrastructure for a large quantum computer. It's a critical component.

Ilia: And the effort is kind of parallel. On one hand, quantum vendors would improve their qubits. They would improve their QPUs, in general, while we improve our interconnect. It's an exponential growth of overall improvement in quantum computational power.

Yuval: As we get close to the end of our conversation, I want to ask you about alternatives. You mentioned that IBM says beyond a thousand qubits, we're going to have to start using interconnects. What kind of interconnects are they thinking about, to the best of your knowledge, if they're not Entangled Network's interconnects?

Aharon: Yeah, so that is an incredibly good question. So IBM, they have superconducting qubits. They work in microwave and are two general options to go outside the fridge. Maybe before that, you can also do a multicore architecture inside the fridge. That is easier. Our software solutions are built to handle that, but you don't need our hardware, which is an optical interconnect inside the fridge, and definitely, IBM is looking into that. But that gets you to about a thousand qubits. Going beyond, you need to go outside the fridge otherwise, you need an insanely large fridge.

Ilia: Which they're building.

Aharon: They're building, but that's going to house a limited number of qubits. And so they have two options. One is the long-term viable option, which is what we're building, an optical interconnect. You convert from microwave into optical, and then you go through an optical fiber to a switch, and it's an optical network. We've been doing these kinds of optical networks for many years. It's a fairly simple device, once you're out, into optics.

The other option is direct communication via microwaves. And that has been shown experimentally where you have a superconducting coax cable or a super-cold coax cable. It's at 15 millikelvins, and you transmit microwaves through that to get one device to talk to the other. That is a much more bulky solution. Imagine these super-cold coax cables going between a thousand different machines, each one holding a thousand qubits to get to a million. It might work as a short-term solution when you're connecting five fridges, but in the long term, we see the only option as being optical.

Yuval: Excellent. How can people get in touch with you to learn more about your work?

Aharon: Yeah, so through our website. We're getting a lot of interest through our website. We make sure to answer every single request that we get. I think that's the easiest. But you can email us, Aharon@entanglednetworks.com or Ilia@entanglednetworks.com.

Ilia: Or Twitter, LinkedIn.

Aharon: Yeah, the usual methods to get people.

Yuval: Excellent. Well, Ilia, and Aharon, thank you so much for joining me today.

Aharon: Thank you.

Ilia: Thank you.

Yuval Boger is an executive working at the intersection of quantum technology and business. Known as the “Superposition Guy” as well as the original “Qubit Guy,” he can be reached on LinkedIn or at this email.

December 26, 2022