Tal David, CEO and co-founder of Quantum Art, is my guest on this week’s Superposition Guy’s Podcast. We delve into Quantum Art’s focus on full-stack quantum computing using trapped ions. Tal and I explore their unique architecture, which emphasizes sophisticated multi-qubit gate operations, the use of optical tweezers for segmenting ion chains, and dynamic reconfigurability to scale up to millions of qubits. Tal outlines the company’s ambitious roadmap, aiming to launch a 50-qubit system by 2025 and a 1,000-qubit system by 2027, with a long-term objective of achieving a million-qubit system. He also discusses the challenges and advantages of their approach, particularly regarding error correction and speed. Additionally, Tal mentions partnerships with companies like NVIDIA and BlueQubit, and stresses the importance of focusing on scalable architectures, among other topics.
Transcript
Yuval Boger: Hello, Tal, and thank you for joining me today.
Tal David: Hi, Yuval. Happy to be here.
Yuval: So who are you and what do you do?
Tal: My name is Tal David. I’m CEO and co-founder of Quantum Art, a quantum computing company out of Israel, working on full-stack quantum computing based on trapped ions. In my background, I have a PhD in quantum physics here in Israel in ultra-cold atoms, an experience in both academia, industry, and government. I used to head the Israel National Quantum Initiative for the government here and for the past three years at Quantum Art.
Yuval: As we know, there are other trapped-ion companies in the market, some of them very big. And so it’s a little bit of Tal David versus the Quantum Goliaths. How are you planning to win?
Tal: Yeah, well, I think that the differentiator between all of the full-step companies is in two major aspects. One is the qubit technology, and the other is the scale up architecture or approach. So we are working with trapped ions. This is the qubit platform or modality that is leading in performance today in the world from the longest coherence times and the best fidelities and up to system-level merits like quantum volume, although I must say that other modalities are catching up and doing amazing work as well. And then the other aspect is on how to scale up the technology and move the systems from the small number of qubits that we have today to the thousands and ultimately millions of qubits that we’ll need in order to get to commercial viability and to fault tolerance in the future. And at quantum art, we have looked at the shortcomings of what other approaches bring to market. e’ve developed our own unique and proprietary scale up solution that looks at these shortcomings and tries to give better approaches on how to scale up quickly and in the best way.
Yuval: I believe you recently published a scaling white paper and perhaps a few months ago also an archive publication about your architecture. Could you explain to our audience a little bit what the uniqueness is in terms of architecture and why that scales?
Tal: Yeah, of course. So this was indeed published in the ArXiv and then on PRX. The architecture stands on four pillars, essentially. First is we look at sophisticated gate operations. We call them multi-qubit gate operations. Whereas the industry, in general, for 40 years, has looked for the minimal gate set, which will be universal, that you can build any algorithm out of, we are working with multiple qubits, with tens of qubits at the same time. Instead of shining laser lights on single or two qubits at the time, and doing sequential operations, we’re operating a sophisticated laser modulation to drive gate operations on tens of qubits at the same time. This gives us a very much more powerful basic gate and saves orders of magnitude in time and also in error. So that’s the first component. The second component is that we need to go to a largernumber of qubits. And in trapped ion qubits, when you go beyond, let’s say, a few tens of qubits, it starts to become a very difficult control problem. Operations become slower and the interaction and the environment becomes harder to control and mitigate. In our architecture, we take very long ion chains, let’s say a thousand ions in the chain and the way that we segment them is not by separating them in space like other people are doing and then shuttling the ions around. Rather, we are separating them by shining optical tweezers and pinpointing in place a couple of ions here and a couple of ions there and in between them, we define segments or cores that now work independently from each other. So you probably know tweezers from other qubit modalities, for example, ultra-cold atoms. But for us, the tweezers are really simple tweezers. All they do is just pinpoint in place a couple of ions in different places and make them into barrier ions, so to speak. And in between these barriers, we have multiple cores working in parallel, so we have a multi-core configuration. And in each of these cores, we can employ the multi-qubit gates that I just described. And then the next step, the third pillar that we have, is how to interact between these cores. ere again, we use the fact that the segmentation is done optically in the first place, meaning that in order to change the configuration and connect cores together, we simply need to turn on and off these tweezing lasers at particular configuration. So in microsecond time, we can reconfigure the architecture or the quantum processor and connect multiple qubits to multiple qubits all along the multi-core configuration. So we can have in a single step, in microsecond time, hundreds of cross-core connections between qubits in a dynamical way. This is kind of reminiscent of how FPGAs works in electronics, that you can reconfigure the hardware configuration, as you speak. And then after we have the three pillars in place, now we have a basic building block, which is not just a two-qubit gate, but rather a thousand ion system with a multi-core configuration and multi-qubit gates. And then we can go and scale it up even further and move from a one-dimensional line into dense two-dimensional arrays of tens of thousands of qubits in a very small footprint of let’s say 10 by 10 millimeters. And ultimately, at the end of our roadmap, we’re going up to a million qubits in a two-inch by two-inch footprint, which is one of the most dense quantum processors out there. his has important implications of being able to reach very large numbers of qubits in a small footprint, meaning that we don’t need to go outside of the vacuum system and connect different systems together, which is a technological challenge and also results in larger and more power-hungry systems. Here we have a roadmap to go up to millions of qubits in a very small footprint. So these are the four pillars of our architecture.
Yuval: When you mentioned multi-qubit gates, so first can you give an example? Is that like a multi-controlled Toffoly gate, or is it something else that you’re referring to?
Tal: Yeah, so basically the basic gate in trapped ions is called the Molmer Sorensen Gate, which is a bichromatic drive that drives a motional mode to couple the internal energy levels of two distant qubits. In ions, because they are all part of the same chain, they are all using the same databus of motional modes to drive the interaction, we have natively all-to-all connected systems. So if I want to do a gate between neighboring qubits or distant qubits, it’s exactly the same effort. What we do in multi-qubit gates, essentially, is we generalize the Molmer Sorensen gate that people know for many, many years in trapped ions, and we are marking what are all of the possible pairs that we have in our registers. For example, if we have a 50-qubit register and you try to map out how many different possible pairs are there, there will be about 1,200 possible pairs. And what we are doing is instead of sending a simple bichromatic laser drive, which is needed to drive a two-qubit Molmer Sorensen, we’re driving a more sophisticated, multi-tone, multi-mode drive to tens of qubits at once. And essentially, we’re doing the same Molmer Sorensen gate, but we’re doing all of them, all of the pairs in one go. Okay, so this is roughly the same thing that everybody’s for many years, but doing that in all of the different possible pairs, instead of sequentially, we’re doing them in one single operation.
Yuval: Does that require special software? I mean, I think there’s a lot of discussion about software portability. I want to run software on one modality and then move it to another. These special capabilities, what does it require from a user from a software perspective to take advantage of them?
Tal: Yeah, so that’s a great question, Yuval. Thank you. I think it really depends on the user or the developer. We do bring Lego bricks that are different than what people are doing in the industry. No one’s working with such large multi qubit gates and dynamical reconfigurability and so on. But it doesn’t mean that you need to go all the way and develop your application using that. So if you are a general developer, you can work with the same single and two-qubit operation, use Qiskit and design your own circuit. And then when you hand it over to us, in our compiler, we compile and transpile to our basic multi-qubit gates and then implement it on our systems. If you are a much more quantum savvy developer and you want to go deeper, you can use our building blocks and design using our building blocks in mind and get more out of the system on your own. So you could be general purpose, but you can also be dedicated. And in that sense, one of the things that we are doing in collaboration with NVIDIA, we just published that a few weeks ago, is working on bringing our building blocks in our scale-up architecture to their Cuda-Q programming language, essentially going in the future to a place where developers using Cuda-Q will be able to use our building blocks whenever they develop their systems and then come to run on our hardware.
Yuval: What can you tell me about error correction? In this architecture, what kind of error correction codes could you implement? And does the architecture have any special advantages as we go to logical qubits?
Tal: Yeah. So first of all, we work with trapped ion qubits, and because they have an excellent base performance and all-to-all connectivity, already this warrants to have some advantages in error-correcting codes and have a low physical to a logical qubit ratio. But on top of that, because we have the multi-qubit gates, that can actually give us full programmability of which qubit is connected to what qubit at the same time as we drive the gates. And because we have the connectivity between cores, we can go even beyond that. For example, we can use the tweezed ions that are not taking part in the actual logical operation of the computer in order to do mid-circuit measurements in order to drive error correcting protocols. We can have error correcting protocols that utilize the multi-qubit gates to do things more efficiently. And so we have, like other ions people are using, the ability to use not only the surface code, which has limited connectivity, but more advanced code, for example, Tesseract code that quantinuum are using and other ones. And we can go even beyond that in compiling to do things more efficiently, and we are looking into that as well.
Yuval: If I remember correctly from my visit to your lab, you are actually building a system, not just talking about it. Can you tell us what the status of these systems are, and when will customers be able to use it?
Tal: Yeah. So we just recently published our product roadmap that is very aggressive and goes very, very far. Our first product, which we call Montage, will be a 50-qubit physical-qubit system. This will be launched towards the end of 2025 into the beginning of 2006 and this will be our starting point and catching up to the state of the art in the industry which today is at about a 50-quit level for trapped ions. Our next generation system, which we are already building in the lab and will be introduced commercially in 2027 is called Perspective and this will be a thousand qubit system up to 100 logical qubit system and this will bring us to initial commercial viability and quantum advantage in the next couple of years. Both of these systems are still one-dimensional arrays using our multi-qubit gates and multi-core configuration. And going beyond that to the later part of a roadmap, in the next few years, we’re going to move from these single one-dimensional arrays into two-dimensional dense arrays, as I mentioned before. And this will bring us to tens of thousands and ultimately to a million qubits in 2033. And of course, employing all of the goodies that we had so far. So today we’re working with simulators. At the end of the year, we’ll start using our hardware systems over the cloud, and then we’ll get to quantum advantage in a couple of years.
Yuval: And when you mentioned 100 logical qubits in 2027, is there a target error rate that you’re referring to?
Tal: Yeah, for sure. But one of the things maybe to note before we talk about the logical qubit error rate is what happens in multi-qubit gates and how do you define the qubit error rate, right? because everybody is used to talk about the two-qubit error or two-qubit fidelity. For us, because each multi-qubit gate can be equivalent up to even 1,000 standard two-qubit gates, we are more interested not in the two-qubit gate fidelity, but rather in the circuit fidelity or the multi-qubit gate fidelity. For example, if you have 1,000 gates of 99.99% fidelity, they will be equivalent to one multi-qubit gate of only 90% fidelity. So this is something to be aware of. Right now, the equivalent two qubit gate fidelity that we are aiming for is 10^-4. And when you get to the logical error, this will translate into 10^-6, and ultimately even to better than that in the longer term.
Yuval: One of the complaints that people have in general about trapped ions is speed. Does that also apply in your approach?
Tal: So yes and no. The overall gate time is quite slower than other modalities. And this happens because the mediator to couple qubits together is through actual motion. You need to have these things moving around and interacting with each other. But when you’re doing multi-qubit gates and you’re essentially in one go implementing tens or even hundreds of two-qubit operations, instead of doing them one by one, you save up on orders of magnitude in speed. So essentially, when you look at the circuit time, we can get down to where the superconducting people of guys are located roughly. And so one of the things that our architecture brings to the table is a way to mitigate this disadvantage that ions usually have. Certainly, this is even more so if you take into account that we don’t need to shuttle the ions around. A lot of the time that is wasted on the circuit level in trapped ion experiments typically is that you move the ions around in order to bring them together and do the gate operations on two-qubits at the time. We are not moving the atoms around at all. We’re just reconfiguring the lasers. And so we save up a lot of time, both in the multi-qubit operations and then in the dynamical reconfigurability of the multi-cores.
Yuval: When you spoke about scaling, you certainly spoke about going from 1D to a 2D array, which would allow you a larger number of atoms, larger number of qubits. Is there a concern about laser power? I mean, how do you capture all these atoms with about a million tweezers?
Tal: Yeah, so contrary to ultra-cold atoms, we are not trapping the ions with tweezers at all. The tweezers are used in our architecture only to segment large ion chains into these manageable-sized cores. So we do interact with the ions. We do need to provide, I don’t know, a few tens of micro Watts per ion in order to drive the gate operations. and the tweezers are on the order of tens of mili-Watts, maybe 100 mili-Watts per tweezer. So on the overall, even if we look at the million qubit system, this is not ultra-high power. We do need more powerful lasers than what usually people are doing, but they are commercial and they are available. We’re working with vendors on scaling that power up, but it’s not something that explodes in terms of power requirements as we scale up.
Yuval: You’ve been doing this for a long time between your PhD and government work and now quantum art. What have you learned over the last 12 months that you didn’t know before that as it relates to quantum?
Tal: That’s a good question. I think that we’re getting real. We’re seeing amazing progress. The rate of acceleration of progress is amazing in all of the companies. And now we need to be a lot more refined and focused in defining where you want to go and defining your differentiator from other players in the market in order to stay relevant and to keep up the pace. I think that one of the key understandings, and I think that we see it also in, you know, in investment deals that you see maybe fewer deals but larger ones is that people understand that you need to focus the effort on the architectures that can really go the whole way towards scalability. I think that we see this is getting more and more realistic and closer and thereby we also find, you know, the engineering challenges are getting to be more and more real. This is not anymore a science experiment. And this is something that we need, we are working very, very hard to do.
Yuval: Tell me a bit about the company. How many people do you have right now maybe a little bit about the funding history and plans, if you can share.
Tal: Yeah, so we plan to raise more money. That’s easy. The company has been on the road for a little over three years. It was a spinoff from the group of Professor Roee Ozeri at the Weizmann Institute of Science in Israel, who’s been an ion trapper for 20 years or so. So we’re based on a lot of heritage coming from the Weizmann Institute. Today, we are a team of over 40 people. We are located in Israel right by Weizmann. We have teams as a full-stack company dealing with building the systems, building the ion traps ourselves, all of the lasers and electro-optics, the software, and also theory and algorithms. So we have all of the capabilities under one roof. As I mentioned before with Nvidia, we very much believe in collaboration. We have many other collaborations as well, both in Israel and abroad. Funding-wise, we raised our seed round three years ago. That was a $24 million seed. To that, we added a couple more million dollars in a SAFE agreement about a year and a half ago and about 12.5 million dollars in non-dilutive governmental grants from the Israel Innovation Authority, from the Israel National Quantum Initiative, and also from the Bird Foundation, which is a very nice binational U.S.-Israeli collaboration platform for industries. And we have a great project with a company called BlueQubit in California, working together with them on quantum machine learning algorithms using our multi-qubit gate architecture and implementing that in the area of image processing. Just a short while ago, Blue Qubit published their papers on their quantum machine learning algorithms using standard single and two-qubit gates, and we’re augmenting that and taking that to the next level together with them in an excellent collaboration
Yuval: So as we end our conversation today, I wanted to ask you a hypothetical. If you could have dinner with one of the quantum greats, dead or alive, who would that be?
Tal: Yeah, so you always ask this question and probably you’ve heard all of the possible answers. I think that I’m most intrigued to have dinner with the quantum great that was born this month. I’m looking forward to what quantum will look like in 50 years or 60 years from now, and I think that we can only imagine where this field is going to, and I would be very excited to get a glimpse into what this will look like.
Yuval: In your estimation, by the way, of when quantum will truly be useful, is when?
Tal: Yeah, so I think that truly useful is a very subjective term, right? So I think that we’ll start to see useful things even in the next couple of years, maybe even in the next year. I think people are starting to have real indications that this is getting to be real, and this is only to go on the playing field. And what we’ll do once we’re on the playing field, the sky is the limit. So I think it will happen sooner rather than later. And I think that the places that we’re most excited about are fields that take our advantages into the application level of not only being very local, but to have both short-term and long-term locality and connectivity and so on. And so one of the areas that we’re looking at very, very seriously is on solving large sets of partial differential equations, for example, in computational fluid dynamics, which has serious implications in aerospace and automotive, but that’s just one area, and I’m sure that there will be others very, very soon.
Yuval: Tal, thank you so much for joining you today.
Tal: Thank you, Yuval, for having me. It was fun.