Yonatan Cohen, co-founder and CTO of Quantum Machines is interviewed by Yuval Boger. They explore Quantum Machines’ role in the quantum computing ecosystem, focusing on their development of control electronics for quantum systems, the transition from academic to commercial customers, the partnership with NVIDIA for enhanced computational capabilities. They also discuss Quantum Machines’ contribution to establishing a quantum computing center in Israel, the integration of quantum and classical computing, and much more.
Full transcript
Yuval Boger: Hello Yonatan, and thank you for joining me today.
Yonatan Cohen: Hi, great to be here.
Yuval: So who are you and what are you doing?
Yonatan: So I’m Yonatan, I’m the CTO at Quantum Machines and one of the co-founders of the company. We’re developing the control electronics, which means all the classical hardware that operates a quantum computer.
Yuval: And your customers are? Who are you targeting?
Yonatan: Our customers are everyone who is building quantum computers. Our mission is to accelerate the realization of useful quantum computers. And we want to do it by providing the technology that will allow the people who build the full stack quantum computers to build better quantum computers. And we believe that the control platform, which is all the classical hardware, as well as a lot of software layers that we’re developing today, which is where our expertise is in, is very important in order to allow our customers to move faster towards their goal.
Yuval: A few years ago when I first heard of Quantum Machines, I saw that many of your customers were academic institutions. So a professor trying to create some qubits or trying to create an experiment. And then I believe that over time you started selling to computer manufacturers. But would a computer manufacturer not want to develop their own control electronics because of cost or IP or other issues?
Yonatan: Yeah, first of all, that’s a great observation and I think it’s been actually fascinating to see the shifts in the field. First of all, from academia to the industry, when we started a lot of the action was in academia. Still is, by the way. I believe still about half of our customers are in academia.
And it’s been great to see the shift that’s happened in the last six years. So we’re seeing more and more quantum computing becoming an industry. And a lot of the, I would say that the main developments are now happening in companies rather than in academia, even though still the academia plays an essential, essential role in developing the next generation building blocks of quantum computers.
So we’re really kind of in between the academia and industry and it’s fascinating to be a part of it and seeing the shift. Another shift that we’ve seen, and this is something that we bet on that we started the company, is exactly the shift from doing full vertical integrations to starting to have companies that specialize in the different layers and then integrators or companies that build full stack quantum computers starting to rely more and more on companies like us that specialize on a certain layer of the computer.
And this was really not obvious when we started the company. Actually, we told investors that that’s our kind of prediction and not everybody believed it. But now we’re seeing this happening. We’re really seeing a shift. We’re saying that many of the companies that develop their own controllers, their own control systems, when we just founded the company would not have done so if there was a company like quantum machines. And to that we’re seeing a lot of new companies that want to build quantum computers by us.
Yuval: As far as control electronics are concerned, so I think there’s a part about driving the qubit, so creating pulse sequences that implement certain gates. But these days people are also worried about calibrating the machine. How do you calibrate it once in a while? And certainly this year more than ever about error correction. How do we measure qubits and how do we correct errors? Which of these three areas, the qubit control, the calibration or the error correction do you deal with?
Yonatan: Okay, great. I love your questions because I think they’re really hitting the point. Some of the things that we put most of our thought and effort on and also were really the source of why we started quantum machines. So we obviously deal with all the layers because when you build a controller it has to do all three things. It has to, right? Like we don’t want to have one controller to do your calibrations and then switch the entire controller to a different controller that can do error correction. That’s by the way one of the challenges.
You want to have the same controller and it has to do both very good because otherwise you need to switch a controller and that’s just not impractical. And in fact as we move from simple control to very sophisticated calibrations that also by the way includes things like error mitigations and very fast retuning of parameters, so recalibration of parameters on the fly and then moving towards error correction that’s exactly kind of the sequence of events that make our technology more and more relevant.
And why is that? Because at quantum machines our core technology is about bringing classical compute power into the heart of the control system. Which means that you don’t only do the control and readout measurements from the controller, you could also close the loop. You can measure your qubits and based on that you can do the classical calculations, the processing that’s needed to recalibrate parameters, retune parameters or to calibrate to begin with even parameters very quickly and close the loop between readout, processing and control.
And when you do that you can calibrate much faster, much more efficiently and sometimes also kind of on the fly. We call it embedded calibration or embedded retuning of parameters. And that’s of course also what’s really important for quantum error corrections because that’s where we need to measure our qubits, process the results, decode what are the errors that we predicted happened in the quantum processor and feed back into the quantum processor from the control system.
And that’s our core technologies. That’s where we started the company. Actually one of our co-founders, Nissim Orfek, he performed the first quantum error correction demonstration on superconducting qubits, which was an important paper in 2016 from the Rob Shulkov group in Yale. And that’s really the kind of history of the company. So we come from error correction and the need to bring more and more compute power very close to the qubits in order to do all of these advanced control and calibrations. And error mitigation techniques.
Yuval: I want to go back to error correction and compute power, but first add another layer if I may. And that is the issue of scaling. So if I’m an academic researcher and I have five qubits, then I need I guess five channels or 10 channels or 15 channels to drive these qubits. But what happens when I have 500 qubits or 5,000 qubits? How does your system scale to support the needs of large scale computers?
Yonatan: Yeah, again, that’s a great point. And in fact, that is exactly kind of so in quantum machines, we had kind of two phases. The first phase with our first generation product, the OPX, that again brought this classical compute power into the heart of the controller. But in some sense, we developed it to work to be scalable only to a certain extent. And maybe on the level of a couple of tens of qubits or maybe a little bit or a few tens of qubits.
And now in the last couple of years, we’ve made the next big leap with our next generation or now it’s our generation, but it’s just been introduced to the market recently, which is called the OPX1000. And that’s a massively scaled-up system that can support controlling thousands of channels. That’s why we call it OPX1000 and thousands of qubits in terms of control readout, in terms of the classical feedback that we can close.
And now the missing piece of it is how do you do the error decoding at such scale? As you mentioned, when you want to do the error correction at such scale, you also need a very, very strong decoder. Now the classical processing power that we brought into our controller with what we call our pulse processing unit is great, but it’s not enough.
And that’s why what we’re doing today with NVIDIA actually, we partnered with NVIDIA to create a very low latency link between our pulse processing unit in the control system and the NVIDIA Grace Hopper platform. And that way, basically we extend our technology so that you can now run very heavy classical processing very, very close to the qubits.
And you can do it all still by programming very easily your control sequences, calibration sequences, and error correction sequences. That’s a key point because it’s not enough that we make the hardware such that it can do all these things. We have to create flexibility to the user to keep tweaking and changing the algorithms, the error correction decoders, the control sequences, the calibrations, etc. from software so that they can iterate over their applications or developments very, very quickly.
So all in all, I think that this is exactly what we’re trying to address now with the built scalable control system and the connection to the GPU CPU platform that can decode also at scale.
Yuval: Putting the flexibility aside, which I know is sort of a big deal in error correction, is GPU the right answer or should I look more at an FPGA or an ASIC solution to do true multi-channel super high-speed simultaneous control of all these qubits?
Yonatan: So the control of one of these qubits is happening from an FPGA, but an FPGA that we programmed in advance so that you can program it via software. So we’re separating the control and readout from the decoder here. And essentially building the right hardware so that you do the specialized thing like the control and readout on your FPGAs. You do the decoding on a GPU. And in any case, you program everything from a very flexible software interface and that is key for us.
And yeah, I know a lot of people are talking about starting to build ASIC solutions and that is important, that is going to happen. But the question is exactly when is the right time to do that because you have to keep the flexibility for the user to keep changing the protocols that they’re running. We don’t know exactly what are the decoders we’re going to run.
So if you hard code your decoder in an ASIC or even on an FPGA and you want to iterate over this, this will take you a month for every iteration. But if you built a platform that gives you the right hardware performance, real-time performance with the right latencies, but you keep it highly programmable, you’ve changed the game. And that’s what we’re trying to do.
And a GPU at the end of the day is the best ASIC that you have in a sense. Someone made a chip that is really, really high performance. Now the question is whether you can connect it to the control units, the FPGA-based control units in a way that you don’t compromise on.
Yuval: There are multiple qubit modalities and surely they have different requirements. How different is the solution that you provide for say superconducting versus neutral atoms or some other modalities?
Yonatan: Yeah, so we’re expanding this as we go all the time. One of the reasons why we’ve created the OPX1000 as a modular platform, which means that you have a box and inside of it you have modules. You can plug them in and out to keep our flexibility to develop different analog front-end modules. We call them FEMs, front-end modules, to become more and more specific as we go.
Fortunately, it seems that a lot of the requirements, very surprisingly in some sense, are similar. Like if you look at what you do when you control flux lines in a superconducting qubit-based common processor, you need the same sampling rates of your digital to analog converters as you do when you control acoustopic modulators for neutral atoms. It’s rather surprising. Maybe it’s because this is the electronics that people have built and everybody kind of converged there, but in some sense it’s surprising.
We kind of enjoy that in a sense that we don’t have to specialize too much, but we do need to develop different modules for different modalities. We believe that one of our values is to do the work that’s needed for different qubit platforms to bring the full value and work with everybody and see how the needs are changing the market as we go.
Yuval: I believe that Quantum Machines has won a contract and is building a quantum computing center in Israel. How is that related to your main line of business?
Yonatan: We’re doing it for two main reasons. One is purely patriarchal. The Israeli government wanted to establish a quantum computing center in Israel to build an ecosystem here in a community. And we believe that we have the right vision for how to do this and how to contribute to this. And we convinced the Israeli government that our vision is the right one to go with.
The other reason is of course that this center for quantum machines is a great opportunity to display our products as a part of a full system. Just like, for example, sometimes NVIDIA will show a full HPC working in order to display how their architecture and their GPUs are doing great in the full system. And that’s very important for us because this is really demonstrating how our technology gives benefits to someone that builds a full stack quantum computer.
The other thing is that it allowed us to really partner with a lot of the important players that are positioned in other layers of the stack and see how our different solutions integrate together. Because we really believe in building this kind of ecosystem where we can start to partner and plug our solutions to others and see how everything works together. And we learn a lot of inputs and different requirements from doing that.
Yuval: A lot of quantum algorithms will probably be hybrid algorithms where they execute partially on a classical computer and partially on a quantum computer. Do you support that or do your products help with doing these hybrid algorithms or is it more that you’re just focused on controlling the quantum stuff?
Yonatan: Yes, so our control system exactly supports this in a very unique manner. In fact, again, we were the first to really bring general classical compute engines into the heart of the control system, which means that we really were the first to deeply integrate quantum and classical. In the sense that in the same program that you write in our programming language called “Qua”, you actually write quantum operations and classical operations in the same code. So you can run a single program with quantum operations and classical operations and with very, very low latency between them. This is the heart of what we do. This is our core technology.
And again, it’s relevant all the way from doing sophisticated calibrations to doing quantum error correction and of course hybrid algorithms, which are very important as the field starts to move into applications. And this is actually also, by the way, related a bit to the Israeli quantum computing center that we’re establishing here in Israel because as a next step, besides again bringing classical compute engines into the controller, we also need to integrate with more general compute infrastructure with the data set.
And that is also done through the control layer. So part of what we’re trying to demonstrate here at the quantum computing center in Israel is how we integrate quantum computers into the classical data center, classical HPC infrastructure, for example, in a way that will allow for such hybrid algorithms to run efficiently. And that’s also where the integration with the NVIDIA GPUs is also a key.
Yuval: If I buy, I’m a researcher, I buy one of your products or several of your products, what can I expect to get other than the product delivered? What does installation look like? What does support look like? Is that a big issue for customers in your mind?
Yonatan: Yes, I think that in our field, one of the most important things that we can deliver to our customers besides, of course, good products with advanced technology is our support. And we actually put it under what we call customer success. We have a very big team of physicists. Many of them are PhD physicists, super experts in quantum computing in different qubit platforms, by the way. And they work very, very closely with our customers. I think that’s one of the key things that we managed to do is when we give you the product, we also give you a really great expertise in how to use the product in your own context, which is very, very important to our customers. And I believe this is one of our key differentiators.
Yuval: As co-founder of a company, what keeps you up at night professionally? What keeps me up at night?
Yonatan: Wow. Companies are a huge roller coaster. Every day there’s like five great things and five things that make you worry. So it’s very hard to answer what’s a single thing, but I’m very, very interested to see how fast we can do something very, very useful with a quantum computer. I think that as quantum machines today, we’re positioned very well in the quantum computing market. As the quantum computing industry, what we need to really focus on is how do we build a roadmap? It doesn’t have to happen in one year or two years or five years. How do we build a roadmap for the entire field that will allow us to bring huge business value at the end of the day to end users? I think this is critical. Of course, this is not our direct business because we’re not selling to the end users. But this is where the entire field is going to derive its value from. And so when I think about quantum, when and how does quantum computing really explode in a good way? I think, you know, what are the killer apps that we’re going to run that are going to really deliver business fun?
Yuval: And related to that, you’ve been in quantum business for a number of years now. What do you know today that you didn’t know six months ago about the industry? What’s new in your understanding of the market or industry? Only six months, not six years. Could be nine months, but not six years.
Yonatan: I am actually more optimistic that we can build a large scale fault tolerant quantum computer in the next decade than I was nine months ago. This actually in some sense surprises me because I’m in the field and the more you are in the details, the more you understand the challenges. But I’ve seen, I think in the last nine months, a lot of very, very interesting progress from many different players. And that made me become more optimistic than we can build a fault tolerant, sorry, large scale quantum computer in the next decade. And I think, you know, being a part of it is really, really, really exciting.
Yuval: And last, as we start thinking about dinner, I want to ask you a hypothetical. If you could have dinner with one of the quantum greats dead or alive, what would that be?
Yonatan: Wow. It has to be either Peter Schor or David Deutsch. Why? It would be extremely hard to choose. Maybe I can take both of them to dinner.
Yuval: Very good. Yonatan, thank you so much for joining me today.
Yonatan: Thank you, Yuval.