Podcast with Murray Thom, Vice President of Quantum Business Innovation at D-Wave Systems – Moving Quantum to Production

Podcast with Murray Thom, Vice President of Quantum Business Innovation at D-Wave Systems – Moving Quantum to Production

Murray Thom, Vice President of Quantum Business Innovation at D-Wave Systems is interviewed by Yuval Boger. Murray and Yuval discuss D-Wave case studies, why customers move or don’t move to production, their future gate-based machine, the applicability of annealers to machine learning, quantum-inspired solutions, and much more.

Transcript

Yuval: Hello, Murray. Thanks for joining me today.

Murray: Hi, Yuval. Great to be here. 

Yuval: Great to have you. So who are you and what do you do?

Murray: My name is Murray Thom. I’m the Vice President of Quantum Business Innovation at D-Wave Systems. So my primary responsibility is to help our customers to build business value or mission value using our quantum computing technology. And I work for D-Wave Systems. I’ve been with D-Wave for a little over 20 years. I started out in the early days, you know, designing for and building quantum computers, and I’ve progressed all the way up to software and working directly with customers just to make that process easier and to help provide impact.

Yuval: 20 years, wow. So I gather you like the company, right?

Murray: Oh yeah, I mean, solving interesting challenges and working with great people and great customers, it’s fantastic. Yeah, absolutely. 

Yuval: One of the questions that often comes up when I speak with analysts is, when will quantum computing be useful? And I think with D-Wave, perhaps you’re saying it is useful now. So what’s your view on that question?

Murray: That’s 100%. What we’ll say is that, you know, we have examples of folks using our quantum computers to support real-world applications. And probably they just haven’t heard of those examples yet. So for instance, the Port of Los Angeles, working with SavantX, have been able to make use of our quantum computers to help them move 60% more cargo per day by optimizing the movement of those cranes, those really large cranes they have at ports, the ones that are called rubber-tired gantries, optimizing their movement to sort of pick individual cargo containers and decide what’s the sequence that allows them to get the most productivity in that day while optimizing, based on the real conditions of which trucks are available to move those cargo containers. 

And then, you know, another example I really enjoy sharing with folks is that Pattison Food Group, which manages 11 you know, grocery chains. And one of those chains that Savon Foods has basically used our technology to help them solve some important grocery optimization challenges, taking them from 25 hours down to two minutes, you know, and it’s gotten to the point now where they’ve expanded into a second application. and they have been in production since October. So when people realize, oh, you know, like in the case of the Port of Los Angeles, when that application was running in production, it was calling our quantum computers every 15 seconds, two shifts a day, six days a week. Then they’re starting to realize, okay, I’m not thinking about lasers on an optics table anymore. I’m thinking about a real compute infrastructure, part of our IT infrastructure that we can access through APIs in the cloud and get results that impact our daily lives.

Yuval: So speaking about the port of Los Angeles, I understand that quantum can solve the problem, but how much effort was put into trying to solve that problem classically? And are you in a position to say that quantum actually outperformed the classical optimizer in this case or cases like it? 

Murray: Yeah, well, I think that in many of these real-world applications, folks are dealing with a lot more than just the technology itself. They’re dealing with a full application workflow. And always, you know, we’re focused on bringing our customers return on investment, you know, value when they’re running it. So in every case that we’re talking about where folks are running these in production, that is exactly what they’re seeing. Commercial value, return on investment in terms of what they’re running. And I think that the ultimate litmus test for quantum computing is being able to fit into that. 

I think that, you know, there have been many papers and optimization and research papers in quantum physics, where we’ve been able to demonstrate situations or problems or, you know, precise information about the right places to use quantum computing to provide advantage where we’ve been able to demonstrate that advantage. But it’s important, I think, for people to realize that that is not the industrial application context. That’s the research context. And in the industrial application context, people are looking for a technology solution that includes all the resources that can help them in, you know, as problems are coming at them thick and fast to get high-quality solutions, whether those are quantum and classical. 

And our experiences have taught us that both quantum and classical technologies have a role to play in these production applications. And so I think a big part of what’s differentiating these solutions and the customers who are using our systems is they’re using quantum hybrid solutions, it’s both quantum and classical computers working together in harmony, complementing one another’s strengths to bring those solutions, solutions to them. 

Now, I think a lot of your listeners may think that actually, you know, it may help them to realize kind of that, you know, there are distinct different levels of demonstrations of performance and quantum computing. One is the commercial value that I was just talking to you about. The second one is, you know, this question of quantum advantage, where like, if a quantum computer is a device that uses quantum effects to accelerate calculations, then a quantum advantage should be a demonstration that the quantum effects are providing that advantage, right? And our recent Nature paper is an example of that, where we can actually directly show both theoretically and through experiments that our quantum annealing systems are able to improve solution quality faster than classical methods. But then there’s another tier above that, which is, oh, So classical computers can’t solve this problem. And that’s like a totally different tier of difficulty where you have to then show that it’s not possible for classical computers to solve the problem. And that’s much more difficult. And so I think that those different levels of questions are kind of getting addressed by different communities with different sort of focuses in terms of how they’re going about it.

Yuval: In terms of solution delivery, do you typically work directly with the customer or do you work through a system integrator or someone that also deals with a lot of the classical side?

Murray: I think what’s most important is to make sure that we’re working with the folks who have domain expertise about the problem that they’re trying to tackle. You know, I think every large enterprise is facing a build or buy decision, like should I build a team with quantum expertise that is going to look within my organization, the places where it will be applicable, and then build my solutions for me? Or is it faster and more cost-effective to work with a group of experts who can do this quickly and are already familiar with the technology? Certainly, I find that if there are folks outside your organization who have a lot of experience with the problem domain, like folks who are very close to the problem domain and what are the features that make that difficult, probably most important, but we’ve certainly found, you know, we’ve actually created an entire program around helping organizations who are looking for our help to accelerate them on that quantum journey, our D-Wave launch program. And that’s really, you know, whether they’re starting at, you know, problem discovery, like where are the places in my business where quantum computing can be applicable, or they’re building proof of concept at application scale, or they’re going to limited production release or full production, you know, that’s a program we can use to help them wherever they’re coming in sort of move along that journey more quickly. And that’s what I think has been very effective and a lot of our customers really appreciate being able to engage with programs like that.

Yuval: The last time I looked at your website, you had dozens or maybe hundreds of case studies with different types of customers and different types of optimization applications. I assume not all of them have moved to production. Perhaps many have sort of stayed at the proof of concept level. And I’m wondering why is that the case. Is it just because they were curious about quantum and wanted to try something, but don’t have real budgets to move into production? Is it because the results were not good enough? Is it something else? Where do you see enterprises on the adoption curve? 

Murray: So that’s a great question. I think that our experience has been that once people understood how to use our quantum computers, they actually found that there were applications for them everywhere. So quantum computers are a lot like classical computers from the perspective that, so for instance, some of your listeners might be surprised to know that however a classical computer is trying to solve a problem, it’s always doing the same task underneath, which is addition and multiplication. And you might ask yourself, what do addition and multiplication have to do with iPhones and driverless cars? Well it turns out if you can do a lot of those calculations and you can do them very quickly, it has everything to do with them. But it’s important to realize that classical computers had early applications and they had some applications that materialized late. And quantum computing is going to be the same thing. So as people have investigated, we’ve seen people complete as many as 250 early applications. And that’s to help people find like, oh, OK, well, here are where discrete optimization or discrete sampling tasks are. Here’s the relationship between quantum computing and machine learning. And some of those are going to basically provide impact at a much earlier scale than others. So I would say that sometimes when we’re working with folks, we’ll actually find a really high-quality technique that is really well served by classical computing. And if that’s the case, we want our customers to be able to take that solution there, because that’s what’s going to work most effectively for them in that circumstance. 

However, if they’re working through a problem, and it’s, you know, everyone who’s coming to work with us on a problem is finding it challenging, they’re not particularly satisfied with the solutions they have right now. And when we do that, we’re able to build a formulation that they’re able to run in production. You know, that’s always thrilling, I think, for everybody. And many of them are kind of in that life cycle. You know, at the process for different organizations progresses at different speeds in terms of how to take some of those early demonstrations and bring them into proof of concepts with their real-world data. And they have to navigate those internal change management processes and we try to help them with that.

Yuval: One of the concerns that customers have about quantum, in general, is that you can only solve toy problems in most computers, right? If you’re running Shor’s algorithm, maybe it’s three times five. That’s sort of the most you can do and probably not even that. How large of a problem can you solve today on D-Wave? I think you mentioned the Nature paper a little bit Maybe that’s an example of a very large problem.

Murray: Yeah, absolutely. Well, I think it’s particularly challenging when you think about quantum computing in the most nascent days where we were looking at four-qubit systems and 16-qubit systems because a lot of times we’re thinking about classical systems which have megabytes of data or gigabytes of data, and a byte itself is eight bits. So if you have 16 qubits, you’ve got two quantum bytes, really at a small scale. And so that’s really at the domain of doing device-level research. Our processors, like D-Wave’s quantum processors, are the world’s largest programmable systems in existence today. They have 5,000 qubits with 15-way connectivity, and they’re available in the cloud right now. They can absolutely support some very challenging problems. And there have been a number of research papers published, exploring, you know, what are the types of calculations that these systems can do that classical computers struggle with. And also, you know, how big of a margin can that provide. Is that margin growing? I think one of the interesting things is that the future for people to think about with quantum computing is that no quantum computer is ever going to be large enough to fit your entire application problem directly onto it like in a machine instruction. The only way to use quantum computers in practical applications is going to be in quantum-classical hybrid solution solving. So this is something where you can pose a problem with a million variables, you know, maybe a hundred thousand constraints and a classical computer is going to receive that problem expression, it’s going to be quite large, and it’s going to start an algorithm which is going to start sweeping across it and doing what classical computers do well, which is like running downhill quickly or trying to reduce the size of the problem space that they need to search. But when they start running into some computational walnuts or some areas of the solution space where they’re having trouble getting unlocked, they can actually break those pieces out, pass them to our quantum computers and feed the answers back in. And that’s a very productive workflow from the perspective of making it really easy to formulate problems, really easy to bring quantum computing directly into applications immediately. And also, I think very importantly, making sure that quantum computing development itself is, is shaping itself to serve that workflow best. So actually improving the quantum computers to serve that hybrid context. And that’s important because if you look at our trajectory in the technology we’re developing relative to other people in the industry, you’ll actually see that the pattern of developments of our quantum computers is specifically designed to produce practical benefit more quickly in that hybrid context. So while others in the industry have been trying to optimize for more abstract metrics, they’ve actually been in some cases reducing the interconnectivity of their quantum processors in order to improve themselves relative to those abstract metrics. But what we have found in a hybrid context working on enterprise customers’ real problems is that increasing connectivity is extremely important. So while our early generation quantum processors had six-way interconnectivity, our current generation processors that we’ve had in the cloud for two years now have 15-way interconnectivity, and our next generation systems are moving to 20-way interconnectivity. And then in addition, as you mentioned in that Nature paper we recently published, we were able to demonstrate that we could operate those 5,000 qubit processors coherently. So for those of you who don’t know what I mean by quantum coherence, I mean being able to do a calculation where the calculation is not interrupted by any interaction with the outside world. So not even like a single photon interacting with the quantum processor while it’s doing its calculation. You know, most folks in the industry are trying to get two qubits to interact with a high gate fidelity because that’s what they’re trying to sort of isolate their systems from. Well, we have been able to demonstrate we can do that at a 5000 qubit scale on our commercial quantum processors in the cloud while our other customers are using the system. So I think that people are going to be pleasantly surprised about the resources that are available to them now, Yuval, and also the rate with which they’re growing.

Yuval: Assuming customers are not in love with one particular technology but just want to solve a problem, so they could say, okay we’re going to go to D-Wave and use a quantum annealer, they could say we’re going to go to someone else and use a digital annealer, not a quantum annealer, or try to use a gate-based machine. How do you compare all three types and if you want bonus points then you can also do analog computing, analog quantum computing. where could it be?

Murray: Yeah. Okay. Well, I think the key thing, maybe the key thing to talk about is the distinction between quantum and classical. So as our quantum computers started to really develop, a lot of effort was, was put into like benchmarking quantum and classical systems, and particularly like, let’s say, our systems which are doing quantum annealing, it’s a, it’s kind of like, imagine thermal annealing is the process of like heating up a metal and then cooling it down really quickly to make it hard or cooling it down really slowly to make it soft. So that’s a thermal annealing process. And you can simulate that with a classical computer, and that’s called simulated annealing. And you can do that to try to help you solve optimization problems. So our quantum computers do that with a quantum process. And it’s the quantum mechanics, which is actually helping you find high-quality solutions. 

So the way we usually talk to our customers about this is that it’s not about quantum versus classical. It’s very, very clear that classical computers are very good at some problems. So if I pass at a problem of searching a landscape and the landscape has the shape of a bowl, no classical computer is ever going to be beaten at solving that problem because it’s extremely good at running downhill. But if I pass at a problem that has much more complex landscapes, we have many papers that have been written now demonstrating that quantum annealers can outperform those quantum computers. So it’s really about why build on a classical-only solution when you can build with technology that’s got both classical and quantum computing in it and is improving with compounding benefits of the developments of both of those technologies. And I think that’s a very easy decision for most businesses to make. 

When it comes to the distinction between, let’s say, circuit model or gate model quantum computing and quantum annealing, I think a lot of folks are not particularly concerned about those distinctions. That’s what they count on us to do. So that’s why D-Wave is building both models is because we’re building both gate model quantum computing and annealing quantum computing because we wanna be our customers’ providers of their full quantum solution. But the key thing I just try to identify for them is that, once you’ve decided to build a quantum computer, there are many different models that you know, at least for how to go about doing that. And they affect, how are the quantum effects used as a resource? What kind of skills do your programmers need in terms of how they’re gonna be programming that system? And what sort of applications do they target? where, like with annealing, the quantum mechanical resource is being used to help the computer move between solutions quickly. You don’t need to be a quantum physicist to program it. So you can actually be a Python programmer, a machine learner, someone in the finance space, or a quantitative analyst. Those are all folks who are readily programming our quantum computers right now. And the last thing is that it’s really gonna be well suited to applications where you have a large number of solutions you need to look through, which is why optimization in scheduling, logistics, transportation, in applications in the life sciences and finances space that are similar, are where it’s really hitting its stride. If you consider the gate model by comparison, then the quantum mechanical resource is allowing you to store more information in your computer. That’s very powerful. It’s also quite fragile, right? Because if the quantum state collapses, the information you put there is lost, which is why you need error correction for that gate model. 

Also as a programmer, you need to know a lot more about quantum physics. So a lot of the software toolkits and SDKs, programming gate model systems are teaching you quantum physics, teaching you about Hamiltonians and complex variable spaces and things like that. And also, we now know that that is not going to be useful for optimization applications, things like the transportation logistics and scheduling tasks that I was talking to you about before. That’s gonna be really in service of applications of like quantum chemistry, simulating electrons and binding and molecules, or fluid flow dynamics, like air flowing over a car, differential equations, those types of tasks. So, they’re serving different applications and that’s why we at D-Wave wanna build both. But in terms of engaging, I think a lot of folks are gonna focus on like, where is the area in my business or in my government mission where I’m not happy with the solutions I have right now, I wanna improve things. And we’re really focused on bringing them examples that say, if that’s the problem you’re trying to solve, here’s the set of technologies that will help serve you the best. And then that’s how they get exposed to quantum computing via the application itself rather than the underlying technology. 

Yuval: You mentioned that you’re working on a gate-based machine and that was obviously in the news a while ago. How’s that going? So just between us, when will we have a D-Wave gate-based machine that we can use? 

Murray: That’s a great question, Yuval. So we don’t have a published timeline for that, But we do have a published roadmap approach in terms of how we’re focusing on the core devices, putting them together in logical elements, and then building up the building the componentry for the quantum processor in terms of how that logic is going to be managed. And, and sort of made available to customers as a, as a computing tool, let’s say, again, rather than just like the core devices themselves. And I’ll say it’s challenging. I mean, there’s a reason why we put a lot of focus in quantum annealing is because we want to be able to bring people practical results in quantum computing as quickly as possible. And with the gate model approach, it’s probably seven or more years before we’re going to be able to help our customers provide impact on their commercial applications with those systems. But it’s a worthy challenge and one that we’re really interested in. And I think a lot of our customers are really interested in seeing, seeing us all further. 

Yuval: We spoke a lot about optimization. I’m wondering if there’s any quantum machine learning application for the annealers.

Murray: Yeah, absolutely. In fact, there’s a very close connection between the underlying programming model of our quantum computers and quantum dealers and machine learning and neural networks from the perspective that like, you know, the qubits are either sort of in a zero state or in a one state in the same way that neurons either don’t fire or they fire. And the qubits are coupled to one another. They actually influence one another’s state in the same way that neurons have synaptic connections between them that sort of determine what patterns of neurons are going to be firing. And it’s that inter-coupling that causes these patterns of events or, you know, patterns or configurations of state and neural networks that make them so powerful in terms of being able to train them and learn on complex problems. And that’s very, very close to the underlying programming model for the quantum processor. 

Now, most machine learning workflows include optimization tasks along them. So there are, you know, if you think about a spectrum of things that are like very near term available to things that are like cutting edge research, then being able to use our quantum processors to do optimization tasks like feature selection as part of a workflow for, you know, building a predictive model of machine learning for or either a classifier or a recommender system, you can basically use our Scikit-learn plugins to start using our hybrid solvers for that future selection right now. Then as you start moving towards reinforcement learning tasks, you’re starting to get more into the research space and developing a lot of your own methods and pushing the boundaries of what’s been done there. And then if you go furthest in that direction into generative machine learning, then you’re starting to look at some papers that have been written on how to take variational autoencoder and then include like a discrete model into to make it a discrete variational autoencoder. So that latent space can actually have more representational power as you’re trying to learn patterns in your data. And then there have been papers written on making that discrete model a quantum model. So you have quantum variational autoencoders. And there are fascinating underpinning connections, technical connections between these spaces, quantum, and AI. And I think it’s just a matter of having these really smart communities have time to kind of like push those boundaries, learn those connections, and program them in that way to show us what can be trained and learned there.

Yuval: As we get close to the end of our conversation, I wanted to ask you about access models. I think early in your history, you sold a machine and delivered it on-site to a customer. Then for a while, you had the machines on a public cloud. And I believe now the access is primarily directly through D-Wave. Can you walk me through sort of the thinking of how this changed over time? 

Murray: Yeah, for sure. I mean, I think in the really early days, we had customers like Lockheed Martin who were interested in having an on-premise solution. So in the early days, that was the service we provided was, you know, setting up and installing an on-premise system. And we did that for Lockheed, and we did that for Google and NASA, and we did that for Los Alamos. 

But really, the growth in cloud access for compute infrastructure was the best thing to happen to quantum computing, because quantum computing is a situation where you have a problem, which is very hard to compute, but it’s not particularly large to express the problem or the solution. So if the problem expression is compact, and sending it to a remote resource is actually very effective, then you’ve got some very powerful resource that’s searching the solution space for you and finding the best possible answers, or in this case, combinations of resources. And then the answers themselves are quite compact. So sending them back is very easy for the user. And by building the lead quantum cloud service, we really opened up access to a much larger community. People got started and got engaged really quickly to the benefit of those commercial applications like the optimizations at the Port of Los Angeles that I was talking about earlier. So that’s really enabling folks to kind of find that their interaction with quantum computing is as simple as any other IT infrastructure they might be using to build applications with, which I think is really important, is taking these complex problems and making them much simpler and much more addressable for our customers. 

However, we’re still seeing interest, I think, for on-premise systems. And we have a system that continues to be in Los Angeles in Marina del Rey with the Lockheed Martin University of Southern California Information Sciences Institute quantum computing center there. And also at Eulog Germany, there’s we have a system installed there. So we actually have sort of a global network of quantum computers that people can access that we’ve been able to install and operate remotely for many years now. And so I think that that’s going to help lots of folks get engaged and get started quickly, that’s going to sort of broaden out the applicability and the sets of applications we’re going to see people using them with. And then we’re going to continue to see, I think, a lot of folks thinking really strategically from the perspective of, you know, wanting to use resources that are in their region, wanting to create programs locally around quantum computing centers that, you know, train the workforce of the future, the kind of of programming skills they need to understand what quantum computers are and what kind of problems they’re solving, how to put them to use. So I think there’s going to be a vibrant market, I think, for on-premise systems as well. 

Yuval: And last, hypothetical, if you could have dinner with one of the quantum greats, who would that person be, dead or alive? Okay.

Murray: Oh, I think for sure that would have to be Richard Feynman. You know, as someone who popularized the idea of quantum computing, you know, and the notion that if we want to be able to simulate these more complex systems, quantum systems, we were just never going to get there with classical computing, and we needed to build, you know, quantum computers in order to be able to address that. And the interesting thing is that when you go back to those early lectures, you know, I can see how he’s, you know, Richard Feynman was a very talented theorist, but he had an interest in the physical implementations and the actual objects that were quantum mechanical themselves. And so he kind of knew that this was not going to be a system that was going to be like a million objects that were all connected. His lattice works of spins and other things were kind of the way that he described his imaginings of how these were going to emerge. And that turned out to be exactly the case with our quantum annealers. So I think it would be fascinating to go back and then kind of reconnect that loop and talk with him about it. But unfortunately, yeah, that’s not going to be possible.

Yuval: Murray, thank you so much for joining me today.

Murray: It’s my pleasure. Thanks so much, Yuval.

Yuval Boger is the chief marketing officer for QuEra, a leader in neutral atom quantum computers. Known as the “Superposition Guy” as well as the original “Qubit Guy,” he can be reached on LinkedIn or at this email.

August 21, 2023

Leave a Reply

Your email address will not be published. Required fields are marked *