D-Wave CEO Alan Baratz on the Path to Profitable Quantum Computing | QBTS Interview

Pleasure to be here. Thanks for taking the time to chat.

Yeah. So annealing quantum computing and gate model quantum computing are two very different approaches to quantum computing that are also quite complementary in the sense that they are each good at solving different classes of problems. Annealing quantum computers are excellent at solving business optimization problems. These would be things like workforce automation or manufacturing plant floor optimization. They are not very good at solving differential equations. That's the domain of gate model quantum computers. They're very good at solving differential equations for applications like quantum chemistry or materials design, but they're not very good at solving optimization problems. So both of these approaches to quantum computing are actually required to address the full set of quantum use cases. Right now, D-Wave is the only company in the world pursuing both approaches to quantum computing to be able to address the full set of quantum use cases.

Yeah, so we started with annealing quantum because it's a much easier technology to work with. It's much easier to scale. That's why our annealing quantum computers are now at over 4,500 qubits, whereas the largest gate model systems are maybe 100 qubits, and they are much less sensitive to noise and errors. In fact, they can deliver good if not optimal solutions without the need for error correction. That's not true with gate model quantum computers. There's no evidence at all that a gate model system will be able to solve any useful problem without error correction. Gate model systems require error correction, annealing systems do not.

No, I think that's an accurate characterization. In fact, our quantum computers are solving real-world problems today. We have customers, in fact, that are using us today as part of their business operations, and our annealing quantum computers are the only quantum computers that have actually been able to solve a useful problem that cannot be solved classically, what we call true quantum supremacy. This was a problem in the area of computing properties of materials that we were able to solve in about 20 minutes, and it would take nearly a million years on the fastest supercomputers. Annealing quantum computers are commercial today. They are solving practical problems. They are being used by customers as a part of their business operations. Gate model systems are still years away from achieving that goal. The work going on with gate model systems today is all in the area of research experimentation, trying to solve small problems or toy problems on the gate model systems to prove the eventual viability of those systems. But we're still several years away from commercially viable gate model systems as we continue to work on the error correction that's required in those systems.

Sure.

Sure. So let's start with what is error correction. One of the things that we all take for granted today is that our classical computers, the things that sit on our desktop, the things we hold in our hand, they all have error correction built into them. We don't think about it, but they're susceptible to errors also. A bit might flip, generating an incorrect answer. That does happen in these classical computers. It's just that we have error-correcting codes that catch that and correct it. Now the good news about classical computers is there aren't too many sources of errors, and they don't happen very often. So the error correction is very simple and very efficient. However, gate model quantum computers have many more sources of errors, many different types of errors, and they occur very frequently. In fact, after about 10 or 20 computations, the errors have built up to the point where your computation is essentially corrupted. That's why we say, without error correction, it's unlikely that a gate model system will ever be able to do anything useful because you need thousands or tens of thousands or hundreds of thousands, in some cases, millions of computations to solve a problem. Yet today, 10 or 20 computations and the computation is destroyed because of errors. So we need error correction within the gate model quantum computers. Well, the codes for correcting these errors are fairly complex because you have to deal with, as I said, many different types of errors. But there's one error that is most prevalent. It's called an erasure error. An erasure error is where essentially the information in the qubit just disappears. That is, in fact, the most prevalent error in gate model systems. About 90% of the errors that occur are those erasure errors. The information just disappears. Well, if you had a really efficient mechanism for detecting the disappearance of the information, that would make it much simpler to do the error correction because you'd only have to worry about the 10%, less common errors that occur. That's exactly what the technology that we acquired when we acquired Quantum Circuits is all about. The qubits that Quantum Circuits has developed, actually, this was technology developed at Yale University and then moved into Quantum Circuits as Quantum Circuits was spun out of Yale. This technology is what's called a dual-rail qubit. You can think of it as each qubit having two wells: one well is for zero, the other well is for one, with a connector in between. Then there's a photon. The photon is in the zero well, it's zero. If the photon is in the one well, it's one. If the photon is in the connector between the two, it's superposition. So what's so cool about the dual-rail qubits is that if the photon disappears, the qubit itself detects that and raises its hand and says, 'I've had an error. I've just had an erasure error.' That's a very easy way of correcting 90% of the errors. Raise your hand, I've lost the photon. Now the system can ignore that data point and move on. Well, what's so interesting about this is that this means that dual-rail qubits can error-correct with far fewer physical qubits per logical qubit. So let's spend a minute on this. There are two main technological approaches to building gate model systems. One is called superconducting, and the other is a variety of different approaches that are not superconducting, like photonics or neutral atom or spin qubits, or trapped ion. Exactly. Here's what's been known up until now. It takes a lot more physical qubits to error-correct a superconducting logical qubit. It can take up to 1,000 physical qubits to get one error-corrected logical qubit with superconducting. It's much more efficient to do it with the other approaches, trapped ion and photonic, neutral atoms, maybe 100 physical qubits per logical qubit. Maybe we can get it down to 25 physical qubits per logical qubit, but the bottom line is, it's much easier to error-correct trapped ion and neutral atom photonic because the qubits are higher fidelity. The qubits themselves are more reliable. They have lower error rates. Whereas with superconducting, it's much harder to error-correct. A thousand physical qubits per logical qubit. Then you might say, well, why would I ever use superconducting? I'll just go with one of these other approaches. The answer is superconducting is a lot faster. Superconducting computation is 1,000 times faster than the compute in the other modalities. So that's why you want superconducting, because who wants a quantum computer that's 1,000 times slower? You wouldn't want that. So you want superconducting, but they're a lot harder to error-correct. Here's the cool thing about the dual-rail qubits that D-Wave acquired from Quantum Circuits. They are superconducting qubits. So they run 1,000 times faster than the other modalities. But because the dual-rail qubit itself can detect 90% of the errors, they have the same fidelity, the same quality as the trapped ion or the neutral atom or the photonic qubit, which means you can error-correct them with 100 physical qubits per logical qubit or less. Now we have qubits that are the best of both worlds. They have the speed of superconducting because they are superconducting, and they have the quality of trapped ion, which means they can be error-corrected easily and efficiently, the best of both worlds. That's very revolutionary and very important. That, in some sense, has just catapulted D-Wave to the forefront in gate-model quantum computing.

Yeah. So the answer to that question is, currently, we get the information after the computation is complete. However, we have in development right now the ability to get and use that information in the middle of a computation. So ultimately, in the quantum state in the middle of a computation.

First of all, I need to provide a little bit of context. D-Wave began working on a gate-model quantum computer several years ago. We acquired Quantum Circuits only in January of this year, but we began working on a gate-model quantum computer several years ago. The path that we were pursuing was fluxonium-based qubits because our annealing quantum computers, our flux-based qubits, we knew them, we understood them, and we believed they had some advantages in the gate-model space. Some of these advantages included, for example, the fact that you could control them with low-speed lines. You didn't need microwave lines. The control was much easier and much less expensive. There were other advantages that they had over transmon qubits as well. But there were a set of advantages that led us to start down the path of flux-based gate-model qubits. Quantum Circuits uses two transmon qubits as part of their dual-rail system. Each dual-rail qubit has two transmons, not fluxonium. Our initial gate-model systems, the eight-qubit dual-rail system that is operational today, the 17-qubit dual-rail system that will be operational later this year, then we're going to 49, then 181, then onto a 1,000. But up through the 181-qubit system, they will use the transmons that were in the dual-rail qubits as QCI developed. What we are working on now is an investigation into whether replacing the transmons in the dual rails with fluxonium would make the control simpler. If so, as we move from 181 to 1,000, we would transition from the transmon to the fluxonium. If not, we'll just stay with the transmon going forward. That's work that we will be doing over the course of the next two years, which is the timeline to get to the 181 dual-rail qubit system.

First of all, the DARPA QBI program, the Quantum Benchmarking Initiative, is an R&D initiative. It's not an application initiative. Right now, it's all about funding R&D to address error correction and scaling of gate-model quantum computers. The reason why it's an R&D initiative is because there are no gate-model quantum computers that can actually solve government problems today. All the funding around gate model is going to R&D. Now, we, with our annealing quantum computers, are able to solve important, hard government problems today. For example, we did some work with Anduril and Davidson, both companies you mentioned a minute ago, on missile defense, where we looked at a 500-incoming-missile simulation. What we needed to do was take out those missiles with countermeasures. Leveraging our quantum computer, what Anduril found was that they could compute the solution 10 times faster. Obviously, when you've got incoming missiles, computing the solution fast matters, and 10% greater takeout rate. On a 500-incoming-missile problem, they could take out 50 more missiles than with the solution they were computing purely classically. This is an example of our quantum computer addressing a real government, Department of War problem today, and delivering strong results. Anduril participated in our customer conference at the end of January. It's called QBTS. They were on stage with me during my keynote, and Anduril's President and Chief Business Officer basically said, 'We were surprised at how capable and mature the D-Wave annealing quantum computer is.' I will tell you that in the roughly month and a half since that conference, we have had many inbound requests from the U.S. government to learn more about our annealing quantum computers because there's now this growing realization that annealing is real today. It's commercial today. It can solve important, hard problems today and deliver value today. The government's, I believe, view of quantum computing is changing as we speak to be more broad-based, looking at leveraging annealing quantum computing today to solve important problems while investing in the R&D for gate-model systems to solve problems in the future that the gate-model systems are more well-suited to solve.

As I said right up front, annealing quantum computers can't do quantum chemistry. In other words, we cannot compute a broad array of properties of different molecular structures. We just don't have the ability to perform those computations. The reason is our qubits in the annealing quantum computer are a bit simpler. They're only spin up or spin down. They're not three-dimensional qubits like gate-model qubits, and that provides some restrictions on the materials and molecular structure work that we can do with the annealing quantum computer. We're excellent at optimization. Gate model can't really address those problems, but quantum chemistry is not our thing. That having been said, AI is now emerging as a very interesting approach to molecular discovery, generative AI for generating new molecular structures. We are able to provide value in the area of generative AI. Let me give you an example. We've been working with a company called Shionogi in Japan. They used to be a part of Japan Tobacco. They were the pharmaceutical division of Japan Tobacco. They developed an AI model for generating molecular structures for new human drugs. Generative AI model. It's a large language model that they trained classically. They then retrained it using our quantum computer. What they found was that the molecular structures generated using the model trained with our quantum computer were far more well-suited to human drugs, basically closer to the dataset than the molecules generated with the classically trained model. Go ahead.

Because the model was more accurate. Leveraging our quantum computer, we were able to train a model that was more accurate in the sense that it generated structures that more closely resembled the training dataset.