Podcast with Paul Lipman - President of quantum computing at ColdQuanta

My guest today is Paul Lipman, President, Quantum Computing at ColdQuanta. Paul and I talk about the differences between cold atom qubits and superconducting qubits, pricing strategies for cloud-based quantum computers and much more.

Listen to additional podcasts here

THE FULL TRANSCRIPT IS BELOW

Yuval: Hello, Paul. And thanks for joining me today.

Paul: Nice to meet you, Yuval. And thank you for having me on your podcast.

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

Paul: I'm Paul Lipman. I'm president of quantum computing at ColdQuanta. So I lead the team, building our quantum computers. And ColdQuanta, as a company, is a leader in cold atom technology. The company was founded in 2007 out of some groundbreaking work that was done at Colorado University in Boulder, where a team, including our co-founder Dana Anderson, were the first group in the world to create a Bose-Einstein condensate. And we can talk a little bit about what that is and why it matters. And so Dana founded the company, and we are one of the world's leaders in developing and manufacturing pristine vacuum chambers in very small footprint. And that's really at the core of what we do that enables a wide array of quantum technology use cases, including obviously quantum computing.

Yuval: Let's focus on quantum computing. What does cold atom mean and why is it different (or why is it better) in your opinion than other modalities of quantum computing?

Paul: Essentially, when we talk about quantum technology and we talk about quantum mechanics and quantum effects, those start to take effect at very small scales and also at very low temperature. So, when you cool the matter down to very low temperatures, the quantum mechanical effects start to get realized, and you can then use them for either creating quantum sensors, for example, as we do at ColdQuanta, and also you can use those cold atoms as qubits. And so, at ColdQuanta, we use a variety of techniques for trapping atoms, cooling them down to microkelvin, hundreds of millionths of a degree above absolute zero, and then those qubits can be used to create a quantum computer. In the case of ColdQuanta, we use cesium atoms as our qubits, and we trap them in a 2D grid of laser light. Then, we use lasers and microwaves to prepare state, to affect the qubits quantum states, to entangle them, and to make measurements.

One of the real benefits of this approach is it's inherently very scalable. Because these are neutral atoms, we can pack them very closely together. There are a couple of microns apart in the array. And so, in a device that could literally sit in the palm of your hand, we're trapping our qubit array. Today, we're working on qubit arrays of roughly a hundred qubits. Very soon we'll scale that up to thousands and ultimately could get to hundreds of thousands, maybe even millions of qubits, again, in a very small space. So, there's some real inherent scale advantages that come with the cold atom approach to quantum computing.

Yuval: So, scale advantages, and also maybe cooling. Do you need that big refrigerator around your computer?

Paul: This is one of the key differences between the cold atom approach to quantum computing and the superconducting approach. In the superconducting approach, you have qubits that are manufactured qubits. They have to be made in a fab and then they have to be cooled down typically to microkelvin to thousands of a degree above absolute zero in these dilution refrigerators. And if you think about scaling up from, say, where some of the superconducting providers are today in the order of 50 qubits to scaling up to ultimately millions of qubits, you have to build these dilution refrigerators that take up an entire room, basketball court sized.

With the cold atom approach we don't require any cryogenic refrigeration. We're simply using lasers to cool these atoms down, essentially to hold the atoms in place, to reduce their motional kinetic energy, and thus cool them down to three orders of magnitude cooler. In fact, in our traps today, we're getting to temperatures of the order of five microkelvin. So, five millionths of a degree above absolute zero, a thousand times colder than a superconducting quantum computer, but with no refrigeration at all. And that has some important implications, again, in terms of how you scale these technologies up, how you maintain state, the coherence of the qubits, which all go to the benefit that ultimately you need in terms of driving real algorithm fidelity.

Yuval: Looking at the flip side, if we had here a representative of a company that makes superconducting qubits and quantum computers based on superconducting technology, what would you think they say that the disadvantage of the cold atom approach is?

Paul: Well, look, I think there's a variety of modalities in the industry. There's the superconducting, which was the first out of the gate, no pun intended, in terms of creating quantum computers. There's the trapped ion modality maybe a little way behind. And then cold atom is really the new kid on the block, but ultimately the new kid on the block with decades of research and technology development and capability behind us. And so, I think if you had somebody from the superconducting world, they will point to the fact that they have these quantum computers in the real world available online for customers to use and experience. Well, ColdQuanta will be launching our first quantum computer, it'll be 100 qubit quantum computer named Hilbert after David Hilbert. And we'll be releasing that towards the end of this year. And then ultimately, from there, scaling up very rapidly.

Yuval: How fast is the cycle time of the actual compute? So, if I had a one megahertz classical CPU, then I know that it's about one micron for each cycle. How long is a cycle for a cold quantum computer?

Paul: That's a great question. So, I think there's two pieces to that. And I think one of them, and maybe we'll get into this and talk about the advantages and how these computers are being used today, comparing clock rates between classical computers and quantum computers while interesting is probably not the right way to think about it because ultimately, we're relying on these devices to do very different things. That being said, the physics of Rydberg atoms, which is the technique that we use for entanglement and for gates, supports clock rates in the hundred-megahertz region. So certainly not today, at least, the gigahertz that you would have from a classical computer. But again, I think we have a bit of an apples and oranges comparison there in terms of the type of work that we're giving to a classical computer versus a quantum computer.

Yuval: Once the Hilbert is available, how do you see it deployed? If I don't need this big refrigeration, do I just own one as a company? Does it go on the cloud? Does it go on your cloud? How do you see deployment initially happening?

Paul: Initially we'll be launching Hilbert, as I say, at the end of this year, and that will be on our own cloud and we'll then be launching on one or more of the public cloud services going into 2022 and that computer, Hilbert, and actually the generations that we have planned to come after Hilbert will initially be hosted in our data center in Boulder, Colorado, conceivably in other locations as well. One of the other benefits of the cold atom approach to quantum computing is the potential for reducing the form factor. And so, we have experienced doing this. I was in our Oxford UK office last week where we pioneered some really interesting work in creating photonically integrated sources for cold atom technology. We took something that would typically be an optical bench of approximately one square meter, and we reduced it down to something, again, that you could hold in the palm of your hand.

And the same thing will be true with quantum computing as well. So, with cold atom, the actual qubit array, you could have a million qubits in something the size of your fingernail and actually with plenty of room to spare, these atoms are packed, as I say, very closely together. And so, the roadmap for us going forward, ultimately our vision is that all the optics and all the lasers and all of the electronics get shrunk down to eventually the point where this can become a rack mountable device. And so, if you think about a quantum computer, say 100 000 a million-qubit quantum computer and a couple of 19-inch rack-mountable units, that eventually opens up some really interesting and compelling use cases. I mean, if you think about a quantum computer at the edge of the network, a quantum computer on a satellite, for example, as part of a quantum communication network, these are things that are not even conceivable for these large-scale room-sized devices, but ultimately, this form fact of reduction I think will open up a whole new world of possibility.

Yuval: Absolutely. You mentioned that the computers will initially be available on your cloud and then going a little bit later on some of the public clouds. Are they always physically in Boulder? Or so if I were an AWS subscriber or a Google Quantum or Azure Quantum, would I end up submitting jobs to a computer that's in Boulder or would it be hosted in one of their data centers?

Paul: Yeah, so I think we have to differentiate there between the near term and the mid-term and longer term. So today in the near term, the same will be true for ColdQuanta as it is for all the other quantum computers that are currently commercially available, which is they're hosted in specialized data centers. And while they may be made available through cloud infrastructure and certainly that has terrific benefits, these are devices that are physically in the vendors data center locations. I think we'll see that change over time and it'll change both as a result of some of the players. Microsoft has big investments in photonic, Amazon are working on their own quantum computers. Obviously, Google has developed their own. And then as I say, as the form factors get reduced, certainly for cold atom, we'll have the capability of then deploying those devices within a variety of data center environments, both public cloud, private cloud, hybrid cloud, it opens up a range of different capabilities there.

Yuval: How do you price the usage? Is it by "oh, I'm using the computer for 32 seconds today, and therefore I pay something times 32", or is it by number of operations or the number of qubits? What's the driver for the pricing?

Paul: This is an area of, I think, quite considerable change that's happening in the industry. And certainly if you look at the pricing of the quantum computers that are available in the public clouds today, it is really all over the place in terms of pricing methodology, in terms of pricing structures, and this is something where we're in active conversation with a wide array of potential customers for Hilbert to determine the most appropriate pricing methodology. As I say, Hilbert's launching later this year, we haven't yet publicly announced our pricing. We have some customers who've said to us, "We just want to pay for schedule blocks of time to be able to run our jobs." And others who've said, "Actually, what we want to do longer term is have a fully dedicated quantum computer, but one that you, ColdQuanta, hosts in your data center and everything in between." So, this is an active area of work for us, and we will be publicly announcing the pricing at our launch.

Yuval: There are customers in the quantum world in various stages of commitment to quantum. There are obviously those who are just thinking about it. There are those who are doing various proof of concepts to see if there's a fit, if quantum really can deliver on the promise. And then there are those who say, "Okay, I'm getting ready to move this into production." At what kind of applications do you think that becomes cost effective? I've been speaking with a couple of customers and one of them told me that he was very excited about, in his case, quantum machine learning. And then when he started doing the math, he said, "Oh, this is going to be so horribly expensive that I'll stick with my classical computers for the foreseeable future." Where do you see it being a cost-effective in the production environment?

Paul: This is really tied, in my mind, to the question of quantum advantage. Quantum advantage, as your listeners I'm sure know, is doing something with a quantum computer that delivers either exponential speed up if that's of importance or solving problems that are classically intractable. But I think it's important to distinguish between quantum advantage in that sense versus maybe what we should call quantum benefit, which is where the QPU is an accelerator or an enhancer of and part of a classical workflow process. So, we can be adding the QPU into that classical existing process. So, if we could do that and deliver, say, a doubling in performance or speed or accuracy, I think that would unlock tremendous commercial benefits. So, it's not simply saying, "I want to take a machine learning algorithm and I want to run it using QML on the quantum computer and I'm comparing the cost and the return of A versus B," but rather maybe a more nuanced view, which is to say there were certain parts of the algorithm, certain parts of the workflow that makes sense to hand off to the QPU.

And it's not that we must necessarily wait for the day when the quantum computer will do something that we literally can't do today with classical, but rather can the QPU provide some incremental benefit and thus look at the cost effectiveness in that lens. And so, I think this transition to commercially useful quantum computing, it's not a bit flip, and it's not going to happen at the same time across all use cases. And I think we’re already seeing some early examples of value being delivered in that way. And I do believe that quantum machine learning will be one of those areas where we're going to see near term impact and near-term results. And coincidentally, I recently bought Santanu Ganguly's book, so I was very excited to hear the interview you did with him just recently. So, a personal area of interest of mine.

Yuval: I tend to agree with you that the issue is cost-effectiveness and not just the “oh, I couldn't do it anywhere else”. And to an extent it's like going from a CPU to a GPU, you could run the same code on an AWS instance that's a CPU only or that has GPU. And if I run it on the GPU, maybe I get something that I get faster response time that's beneficial to my customer or otherwise save money. So, given that we're talking about existing applications, just part of them being executed on a QPU, if you were a betting man, is it just QML or do you see other applications that you think are very close to that cost effectiveness threshold?

Paul: That's a great question. And actually, you made me realize there's an additional angle actually to a previous question you asked, which is about where the QPU will be hosted in this public cloud scenario. And I think one of the other factors and the example you gave there of the GPU is part of the machine learning process is really an important one, which is we ultimately, in production, will need to deal with latency effects. And that's the reason and that's the imperative for the QPI to actually be physically co-located with the classical CPU environment. So, I think that's also a consideration that we need to be mindful of. 

Coming back to your question though, in addition to QML, so we're seeing a lot of interest in having a lot of very compelling conversations in the quantum chemistry area. So in terms of molecules, simulation, pharmaceutical, I think we'll see some early work there with customers that we'll announce going into next year. And for us that's really a function of where we can bring 100 qubits to bear and now, we can start to map some pretty complex problems to the QPU. And then also in financial services there's really a lot of interest we're seeing in derivative pricing, in portfolio optimization and I think some interesting results that'll come out of that industry really in the near term.

Yuval: So, you guys are making a lot of progress on the hardware, first version later this year, subsequent versions in the future. Other companies are also racing towards faster or larger quantum computers. Other than the hardware, what else do you see as a barrier to quantum computing going a little bit more mainstream?

Paul: I think this could be the subject of an entire discussion in and of itself. And there's really some very key points there. I would highlight maybe three things. I think first the programming platforms, and I know this is the space that Classiq operates in, traditional developers, software developers don't write and not gates. They're not coding at that detailed gate level. And the same will have to become true for software developers that are writing algorithms for quantum systems. So, there's that abstraction layer that is going to have to emerge and it's going to have to mature. And similarly, I got my start in quantum computing learning to program Qiskit myself. And you could write a circuit for five qubits, maybe 10 qubits, perhaps if you're far smarter than I am, you could scale that up to 50, but there's no way to hand-code gate level circuits for hundreds, much less thousands or even millions of qubits.

So, that whole programming platform operating system level will have to emerge and mature so that these quantum computers and quantum systems can evolve out of the early experimental way they're being used today to really be used at scale in production across a broad swath of industry. I think that's the first point. 

I think the second point ties back to a word you used in the question, which is production. I think to really be effective as part of a workflow process in production, these hardware platforms and systems have to achieve commercial class, availability and performance. If I'm running, again, the derivative pricing system, let's say, for an investment firm or I'm running an optimization process for a logistics company, and part of that workflow is being handed off to the QPU, then that system has to be available 24/7. It has to meet the set of service level metrics. It has to be performant and responsive. And so, I think there is a maturation process that has to happen in the quantum ecosystem to support that. 

I think the third, and maybe this one ultimately long-term is the most important, is the topic of the quantum workforce development and training. And this is across the spectrum, everything from PhDs, we're hiring as fast as we can, AMO physicists and quantum physicists. And so while there is a pipeline of PhD candidates, we certainly benefit from a very close relationship with Boulder and with Madison in terms of the physics departments there. There is a need for an expansion of that end of the pipeline, but this is really all the way down, I think, through high school level, education of quantum technology and its importance and its applications both on the scientific side but also on the business side as well. So the business decision makers in companies can understand what quantum is, why it matters, how it can be applied, how they can derive benefit from it. And so I think that education plays a very major part in this question of what's holding quantum back from more widespread deployment.

Yuval: Perfect. And Paul, I know we're running out of time. How can people get in touch with you to learn more about your work?

Paul: So certainly, our website, coldquanta.com, is the best way to contact us and certainly follow us on Twitter. Connect with us on LinkedIn. We also have a Clubhouse channel, Quantum Revolution, that I would encourage people to check out. And then to connect with me personally, you can connect with me on LinkedIn, it's Paul Lipman, and certainly feel free to contact me by email, it's paul.lipman@coldquanta.com.

Yuval: That's perfect. Thank you so much for joining me today.

Paul: Very nice chatting with you, Yuval.

‍