Podcast with Vishal Chatrath, CEO and co-founder of QuantrolOx

Transcript

Yuval Boger: Hello Vishal and thank you for joining me today.

Vishal Chatrath: Thanks a lot for inviting me, you are very nice to be here. I’m a big fan of your podcast, so I’m really honored to be here.

Yuval: I appreciate you saying that. So who are you and what do you do?

Vishal: So my background is really from the technology side. I entered the tech industry after graduating from India in 1995. My first job as a research scholar was at the National University of Singapore. So, I worked on three-five semiconductors. And in those days, trying to do something that we call the quantum wells among various things. Then in ’97, I thought there’s so much material research wasn’t for me, so I jumped onto the industry. I didn’t know all these years later I would come back to quantum. So I worked in manufacturing in Singapore for a while, then I moved to North Finland, and worked for Nokia for many, many years, in variety of roles, operations, mobile architecture definition very early, business development, corporate strategy that took me to many different places around the world. I lived in six countries, I think about 11 cities by now. And the common thread always was to work with very early-stage technology. And that’s really my comfort zone, something that’s fashionably called Deep Tech today. 

And I’ve been an entrepreneur since 2010 with varying degrees of success and learnings. And my previous two startups were linked to University of Cambridge, both in AI and machine learning before it became fashionable, and started Quantrolox in 2021, I was invited by Professor Andrew Briggs to his team. And Andrew was one of the pioneers of the quantum program in the UK. The other three co-founders with me, Professor Natalia Ares and Dr. Dominic Lennon is also from the University of Oxford. So they are the people with a physics background and I am this so-called industry guy trying to take the company out, take the science to technology and build a viable business. 

Yuval: And what does the company do?

Vishal: What we focus on is trying to develop machine learning-based software that automates the tuning of qubits inside the quantum computer. As we all know right now that qubits are highly, highly unstable as compared to your bits in classical computers and qubits don’t like to stay where we want them to stay. And if you look at your typical superconducting or spin qubit-based quantum system, you require two to three researchers to literally babysit it, you know, kind of just to keep it alive. And we’d rather have these very smart physicists rather than being there, trying to babysit the systems, doing something useful, so our automation software will hopefully bring about complete hands-off, keeping the things alive by themselves. Think of it as an autopilot for quantum computers.

Yuval: How does it work? Does it think about the quantum computer as a black box, just runs some tests as if by where a user and measures the results, or does it do something else? Please, whatever you can explain.

Vishal: Yeah. That’s a very good question. I’ll keep it at a slightly higher level. So what we do is we take the readouts from the quantum computer and at microsecond intervals, we make the decisions which essentially among various things tell the AWG, which is the arbitrary waveform generator, what is the shape of the pulses to send back to the quantum computer. And we are making these decisions at microsecond intervals, so it’s a very fast, closed-loop automation that learns in real-time from the status of the machine. We’ve been testing these ideas and algorithms in the university environment for maybe the last four or five years, quite successfully with spin qubits. And the first product, which we previewed at APS March Meeting and will be launching at SQA, which is the Superconducting Qubits and Algorithms Conference on 29th of August in Munich. and it’ll come integrated with a few different control hardware types, including Qblox, Quantum Machines, Zurich Instruments, and a few others.

Yuval: Given that it’s microsecond resolution, is that a software product or a hardware product?

Vishal: It is a software product, and that’s a very good question that you asked. Our integration is very much close to the actual control hardware, the kind of a lower-level integration with the hardware. And I think that also makes it makes us different from some of the other companies that are trying to do control software, but they do tend to come in from the kind of higher layers like Qiskit  where we are connecting and talking to the control hardware at the FPGA level.

Yuval: Wouldn’t a quantum computer manufacturer develop their own control software and own optimization software or do you feel there’s a need for more of this componentized approach?

Vishal: I think that’s a very good question. Because I love technology and I also love history. So if I look at the birth of any industry, in the absence of a supply chain, in the beginning, everybody builds their own things. So if you look at IBM in the 1970s, there was no notion of, a separate BIOS  OS, application layer, UI layer, and so on and so forth. It was just one, monolithic stack where all these  distinct layers were not there. As only in somewhere around 1984, when IBM came up with the MDS architecture, which started to  differentiate out these layers. The really big ones who can afford it, like the IBM’s and Google’s for all the right reasons, they are doing everything in-house. But there’s this huge market out there of people who are not building their control hardware and are buying it. And I think that’s the market we are going to address to also give our control software because there is just too much to do. And there are many companies who are trying this fully integrated, full architecture approach. But you can already see, unfortunately, from the market cap of the not Googles and the not IBMs, that it is a very expensive game. And people have to open up their architecture to be able to succeed and compete. 

Yuval: Is the technology optimized today or specific to one type of qubits or a couple of types of qubits or is it more universal in nature?

Vishal:  In theory it should be applicable quite widely. But so far we have only tested it for spin qubits, which was our history, and now for superconducting qubits and it works very well for both. Because we are a small startup, we are only 18 people, in the beginning, we will not go out to every market. We are very much trying to build a viable business. Wherever the market volume is, we go after that. So hence the first product is for superconducting qubits because the market volume.

Yuval: Is this essentially a real-time calibration or do you take up a certain percentage of the availability of the computer to run periodic calibration?

Vishal: So the software works in kind of real-time and all the time because I think there is so much happening in the environment.  There are about 40 parameters we are taking into account while trying to keep these qubits alive. And it is real-time, but as an  engineer, you might argue that actually what is real-time. So I can tell you at microsecond intervals, we are making the decisions, and the software just runs all the time to keep it alive.

Yuval: I assume the desire is to scale to hundreds or thousands or more qubits. How can you run this for something with microsecond intervals and have thousands of qubits? You’re going to need a little supercomputer to run the calibration all the time for all the qubits at these kinds of intervals, wouldn’t you?

Vishal: I think yes, we will need a  good enough machine, but at the same time, there will also be quite a few smarts on the software side. So for example, right now, because we don’t have access to any chip that’s over, I would say 20 or more qubits right now, we hope to get bigger and bigger chips.  Even at the scale of a 20 qubit chip, , you have to do some parallelization of the workload.. Because if you try to do it in a serial order, even 20 qubits, the whole tune-up will take a very long time. So our goal is to be able to finish the tune-up from scratch in a few minutes. 

Yuval: For the real-time tune-up, it sounds like you might essentially be tuning the qubit in the middle of the program. While the program is running, if I have multiple operations, then the first operation might have qubit tuned to one parameter and then later on to another one, given, as you mentioned, the microsecond interval. Is that correct?

Vishal: You bring up a good point.  For the period in which the compute is on process, then the goal is not to touch them. But once the compute is over, then to kind of bring it back to the correct state . I think you are very accurate with your description.  Our software ensures that when somebody wants to run a compute,  the qubits are available in the desired state of tune. Because right now, and I’ve been kind of talking to one of the industry leaders, the state of the art is that it’s actually a human doing the tuning. So if they want to run any kind of a ten qubit algorithm,  there’s a human who goes and tags the kind of a ten best qubits inside a QPU, and then they run the compute, and then he or she will shout across the room that okay all is good, and then they will click enter, and the program runs. So, we are trying to automate that entire loop.

Yuval: I think historically computers were calibrated every couple of hours or maybe even every couple of days, and then during the between that and the next calibration cycle, the performance would decrease. You’re doing continuous calibration. In your experience, in your testing, how much better does it perform than calibrating the machine once in a while, even if it were once every five minutes or once every couple of hours?

Vishal: We cannot talk about any results yet, but that’s something we will show at our product launch in September , I think that’s a good time to talk about it. What we have definitely seen is that the calibration is  much faster than a human. And also our first goal is to not focus on the qubit quality in itself because at the end of the day, we can only drive the qubit to the ultimate capability of the qubit hardware itself. Our goal is at the first instance to speed up the process, and once we speed it up the process then work on the quality metrics. So it’s two different parts we are driving. 

Yuval: Some optimization tools understand that not all qubits are the same, some are better than others, maybe based on manufacturing variances or others, and then work to optimize the program to make sure that the best quality comes out. Does your software provide information to allow that software-level optimization of the program itself.?

Vishal: So there are two parts to it. The first part is that our software is very sensitive to the individuality of the qubits, and  we have some research papers on our website.

One of the key papers was where we showed that for three completely different silicon germanium spin qubit architectures, without any changes in the software, we were able to learn the kind of intricacies of the devices and be able to tune them automatically. What we don’t have yet in the first phase is a link between the need of the state of the qubit depending upon the algorithm. There’s also some of work going on,  where for a certain algorithm type, you will be able to bring the qubits to be at a certain state automatically to be able to execute it. That is something for the future.

And while we are working in the first stage for integration towards the control hardware, in the second phase, probably sometime in 2024, we’ll start to work on the integration of our software, upwards to the openQASM  layer. where we want to make the link between the needs of a particular algorithm and to ensure that the qubits are automatically driven to the desired state..

Yuval: And when you say optimizing qubits, I actually assume you mean optimizing gates, right. It’s not just a qubit, but two-qubit gates?

Vishal: You’re absolutely right.  That’s what I mean. In our head, we are so much focused on the qubits itself. And the goal obviously is to automatically get to single-qubit and two-qubit gates to a level as defined by the needs of the higher layers.

Yuval: You mentioned that you use AI and machine learning to do this. Machine learning today is an overused term. Everyone wants to attach AI or machine learning to their product to make it seem smarter than it is. To what extent is it really machine learning as opposed to just a smart, closed-loop system control?

Vishal: Our focus is specifically on Bayesian optimization, and we use Gaussian processes and Bayesian optimization for that. And I have a bit of history with that kind of machine learning and not really deep learning. So the previous company which I founded and I was the CEO for five years was Second Mind, where we used similar techniques also for the tuning of internal combustion engines. So, in that case, it was a 13-parameter state space. And again, because the cost of data acquisition is so high, we simply cannot get the amount of data that you require to use deep neural nets from a quantum computer or from an internal combustion engine. Therefore, the team and myself, and our CTO, we have a very good background in applying those very efficient machine learning approaches. 

Yuval: From combustion engines to qubits, that’s an amazing journey. Do you need special feedback or special data collection inputs from the quantum computing manufacturers? Do you go to manufacture and say, well, look, if you could give me XYZ parameters, we could do much better tuning?

Vishal: Yes, of course, you know, that helps. So I think also as far as the company’s concerned, and especially in this, the environmental, so when you use a Bayesian optimization, the modeling of the environment has to be right, because if you get the wrong model of the environment, you’re going to get all junk. So we’ve been very fortunate that my co-founder Dominic Lennon, when he did his PhD under Natalia’s lab, he had really focused on being able to get these machine-learning techniques and quantum physics in the same brain. So, in that sense, yes, we do have a very good understanding of the actual physics that we are dealing with. And we, of course, work very closely with Quantware and QBlox and Quantum Machines and  Zurich instruments to ensure that we have a deep understanding. And the other thing to remember is that we are all experimentalists. So, all the machine learning techniques that have been developed over the last five years were not done in some simulation environment but actually done inside in Natalia’s lab,  which has five dilution fridges. So we are very much hands-on. So we know what works and actually what doesn’t work in real life. 

Yuval: Some vendors say that they can use the same techniques that they use for qubit optimization or gate optimization to also optimize quantum sensors. Do you see it the same way? Would your approach be applicable to quantum sensors as well?

Vishal: I think it depends upon what you’re using to build the sensors. So these days most of the sensors that I hear about are using, you know, iron traps and , cold atoms and so on and so forth. And for any of these, , laser-based approaches we don’t have experience. Again in principle, it should work, but in practice, we have not tried it. Our experience is with any sort of microwave-based system.

Yuval: What can you tell me about the company in terms of size, funding, what you’re looking for, location, or anything you can share?

Vishal: Yeah, so we started at the University of Oxford, and we have two main locations. One is in Finland, and one is Oxford. We are about 19 people now, almost 50/50 between Finland and Oxford. We have raised over 5 million in terms of venture funding. We’ve gotten over 3 million in terms of various grants from EIC, the UK and Finland. And right now, we’re in a very exciting place where from day one, we’ve been working very closely with the industry. So I believe we have one of the first companies to focus so much on end-user research. So when, for example, we were thinking about the first product, because I come from a product development side, I really wanted to understand that day in the life of a quantum scientist because we knew that part of the improvement that we bring will come from our smarts. But I also observed that any typical quantum scientist, in order to control the quantum computer, was jumping anywhere between 35 to 40 different windows. And to me, that didn’t sound very operationally efficient. And we thought, can we bring in software that  brings everything together and solves the problem down to one window? We really think about a lot of these parameters. So because we’ve been building this product so closely with actual quantum computer users, we believe that when the product is fully released at SQA at the end of August, we are quite confident it will really solve some real problems for quantum scientists. And what we would like, we would love people in research labs, in academia, in companies to really try our software and actually give us feedback. And that’s the only way to build an outstanding product. What we do want to be is we want to be the de facto sort of control software layer. Or at least we want to work towards trying to create some standardized interfaces. And we are again very much keen on this whole open architecture approach, which we believe is foundational to scaling of the industry.

Yuval: As we get close to the end of our conversation, I’m curious, what have you learned over the last six months that you didn’t know prior, what’s new in your world or in your view of the quantum world.

Vishal: Well, first thing is that after dealing quite a bit with dilution fridges, so we have actually realized how fragile these systems can be..  And how much more automation is required. So far we’ve been focusing on the automation of qubit tuning. Eventually, we’ll have to get to the full end-to-end automation of the entire quantum computer. Our focus has also changed, ambition has changed. So now I see ourselves as an end-to-end software for, as I mentioned, controlling the entire quantum computer. The parallel that I can bring in from my world of manufacturing is that I, when I first entered manufacturing, you had about maybe 10, 15 machines in the production line and each machine had its own computer to optimize that machine, but you also had this software that controlled the entire manufacturing line. So I think there is a need for that. And eventually, we see ourselves becoming that, a software to automate the entire end to end operation of a quantum computer

Yuval: And a hypothetical question, if you could have dinner with one of the quantum grades that are alive, who would that be?

Vishal: The thing that really got me excited about quantum physics in the first place, was a book that I found in a bookshop called Surely You’re Joking, Mr. Feynman. And I was a teenager then, and that’s the time I realized that you could do physics and have a lot of fun. And so it has to be him. And I would just talk to him about his life and if he still has time to take me through at least one of the Feynman lectures on quantum mechanics because that’s what I really grew up with.

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

Vishal: All right. Thank you very much, Yuval. It was a joy.

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.

June 5, 2023