Podcast with Olivier Ezratty and Rob Whitney, co-founders of QEI, the Quantum Energy Initiative

Yuval Boger: Hello, Olivier. Hello, Rob. Thanks for joining me today.

Olivier Ezratty: Hello.

Rob Whitney: Hi.

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

Rob: Okay, so thank you very much for inviting us. So, I’m a quantum physicist. I worked on quantum physics since the late nineties, but in the last three or four years, I’ve been really been more interested in the quantum technology aspects, in particular quantum computing, before that it was more fundamental questions of quantum physics and now it’s a bit more towards applications.

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

Olivier: Well, I’m an author of a book you know, Understanding Quantum Technologies. I’m teaching, I’m training various companies, and I’m also doing a so-called techno screening. As part of the work we’re going to talk about, I’m the engineer crossing the chasm between hardware, software, engineering, and enabling technologies.

Yuval: I recently learned that both of you started an initiative called, I think, QEI. Would you mind telling me a little bit about that?

Olivier: Yeah, so the Quantum Energy Initiative was created out of the idea that we need to look at the energetics of quantum technologies very early on. We need to understand it, how it works, how we can optimize it. Particularly with regards to quantum computing, we know that if we don’t address that question, we may end up with having two power consuming systems. We know also that if we work on the optimizations of those system, it could create very powerful and competitive competing systems to the energetic. So, it started actually last year with a PRX paper from our colleague Alexia Auffèves called ‘Quantum Technologies Need a Quantum Energy Initiative’. So, we did it. So, we launched formally an initiative about creating an international community of researchers and industry vendors looking at these issues, working together, sharing some methodologies, sharing also science in an open way.

So, at this point in time, we have a website, we have a community of about 300 participants. We have a board, we just created a scientific board, and we even have some industry vendor partners. So far, we have 24 partners, mostly from Europe, I would say at this point in time, like IQM, Alice & Bob, Pasqal, Parity, QC in Austria, Kilimanjaro in Spain, aQuantum in Spain. We want to expand that, of course. One of the last news from the… I would say, the activity standpoint is, we are launching our first workshop that will happen in November in Singapore this year. We already have a very interesting line of keynote speakers like Peter Zeller from Austria, we have Michelle Devereux from Yale, Janet Anders from Exeter in the UK, and others, even from Japan. So, that’s where we are right now. Yeah…

Yuval: So, help me understand this. I think the conventional wisdom is that quantum computers take less energy than classical computers. So, one of the potential advantages of quantum, beyond, of course, the computational capabilities, is that you can do the same thing for much less energy. If that’s the case, why do we need QEI at all if we believe that quantum will take much less energy than classical?

Rob: Okay, that’s a very good question. In some sense, that brings us to the research that we did in the CNRS to answer this question. Because, on one hand, some people are saying exactly what you just said. On the other hand, some people more or less saying the reverse, if you look at experiments that are done on a few qubits in the prototype quantum computers, now they’re consuming… Sometimes even in the kilowatts per qubit or a bit less, maybe a few tens or hundreds of watts per qubit. So, if you want to scale that up to millions of watts, you immediately start to see power consumptions of multiple nuclear reactors. So, then the question was how do we scale towards large number of qubits without exploding the energy bill? That brings us onto our research topic, which then led us to realize there’s so much more to say on the domain that we need to create a broad interdisciplinary initiative.

So, we just started with small elements, taking the expertise that we had around us here in Grenoble, in the CNRS, and the CEA, and built a simple model of a complete quantum computer from the end user to all of the electronics, cryogenics, everything you can think of. So, the end user means they want to know the solution to a particular problem. So, we took a problem that we believe that most of us believe is not soluble on a traditional computer, which was breaking RSA encryption. Not because we necessarily think it’s the most exciting problem to solve, but because it’s one that’s well documented. So, we took that procedure of taking the best-documented things for every element in the computer, and tried to build a model where you could actually answer this question, where you could optimize all the parameters and actually answer the question, “Well, how much power will it consume?”

So, we are very happy that we managed to build the model, and then we put in various scenarios, and depending which scenario we took, we got all possible answers. From consuming a power consumption of multiple nuclear power stations on the worst cases, to consuming the same power as a few Tesla cars on the best cases. So, now it’s a question. So, now that was just a first scan of the domain. So, now, but already it gives us indications and hints of things to avoid, and things to go towards to avoid going into the regime of multiple nuclear power stations. So, I guess that’s already the first indication. I should say we took a particular technology for our model of superconducting qubits a little bit like vaguely inspired by Google’s computer, Sycamore chip, but with rather futuristic view. So, futuristic improvements.

Even then, we sometimes still got into the regime of multiple nuclear power stations. So, one has a lot of work to do at the level of the enabling technologies, to control electronics, and all of these things if you want to get down to more reasonable power consumptions. That brings us to the interdisciplinary nature that we found that there are so many feedback loops between different scales in what we call the ‘full stack’. So, from how you choose to do the algorithm affects how you want to design the hardware, and how you design the hardware affects how you want to do the algorithm, and how you do the cryogenics affects how you want to do the electronics, and all of these sorts of things.

So, it really shows that you need an interdisciplinary group, which we are trying to construct by creating this community of people who can bring better information about the cryogenics from the cryogenic engineers, better information than we have about the electronics, from the electronic engineers, better information about the algorithms from the quantum information experts, and bring that together and see if we can create a better, more realistic model for more different kinds of quantum computers, which gives us some better guidelines about how to compute without the cost being extravagant, let’s say.

Yuval: Let’s break it down for a second. So, you mentioned superconducting qubits. So, out of the total energy budget of such a computer, what are the key parts? Is that the cryogenics take 95% of the energy, or is it something else? How does it break down for, say, superconductors?

Rob: So, for superconductors, the big expenses seem to be, so every unit of heat that you generate at a cryogenic temperature, you have to spend a lot of electricity to extract. That’s just the law of thermodynamics. So, typically for everything that’s inside the cryogenics, so in at cryogenic temperatures, it will be the cost of keeping things cool, which dominates. But what we also realized is… So, that’s a big cost, and that was what we were expecting, but the surprise was that we realized that the control electronics are a similar order of magnitude for the budget, even when they’re sitting at room temperature. So, all of the electronics that generates all of the microwave pulses that are going to go to the qubits to control the qubits, and to read out the qubits is not negligible, which we weren’t expecting. If you could make your cryogenics quite good, quite efficient, then it could even start to dominate the budget.

Olivier: What we start to see is that it’s very different across different types of qubits.

Other types of qubits will have different energetic builds. Sometimes the lasers will consume a lot typically for trapped ions or for cold atoms. In the case of photons, you’re going to have mix of lasers, of controls also from cryogenics, but less powerful because you don’t cool it very low, you cool it at 4k, you cool the detectors and the photon generators. So, it’s a very different, and it’s very interesting to look at the energetics of all these various technologies, not just one.

Yuval: Does that lead you to recommend one technology over the other, say trapped ions are going to be much better than superconductors or something else?

Rob: So, given that in the superconductors, we have scenarios that vary from nuclear power stations to a few Tesla cars. I think it’s definitely way too early to say that there’s a technology that’s going to win and another technology that’s going to lose this.

Olivier: But there’s some insights still. The insights we have for trapped ions for example, is we know that to scale trapped ions, you will need very early on to connect different QPU’s, so different quantum processing unit. So, then we will have to find a way to scale the interconnect technologies using photons, usually using conversions in some cases. So, we will have to look at the energistics of these interconnect solutions to have an idea on the way to answer to your question. So, that’s requires to have a lot of people working together there as well.

Rob: Yeah, yeah.

Yuval: If we pick on superconducting qubits, I don’t think they want to take more energy on purpose. They say, “Well, we have to have cryogenics, we have to have control electronics,” and so on and so on. So, assuming they’re not switching to a different modality, what can a superconducting vendor do to reduce the energy footprint?

Rob: Okay. So, there are a lot of whole bunch of parameters. So, maybe just to just mention a couple of them and not to try and be exhaustive, it’s probably more interesting. So, the first one that comes to my mind is the wires that come down from room temperature to the qubits conduct heat, and there’s a limit to how much you can reduce the heat conduction of those cables. So, the smart thing to do is to say, “How can I reduce the number of those cables?” That’s all about multiplexing. So, if I can put 10 qubits at the end of one wire and still address them individually, for example, by putting in the different frequencies or something like that so I can talk to all of them individually, then that’s 10 times better than if I have one qubit at the end of each wire.

So, those clever tricks about multiplexing, and reducing the number of wires per qubit are clearly winning strategies, I think. This is one of the things we realize is there’s always a price to pay. So, the price to pay is more crosstalk between qubits. So, you send information to one qubit, but if there are too many qubits being addressed by the same wire, then it will tend to increase the noise that’s felt by the other qubits. So, you need to find the sweet point between the two. That’s where an optimization of the type we’re thinking about enables you to make progress by finding the sweet point between energy consumption and quality of the qubits. Which brings me nicely to one other thing I wanted to mention, which is for us we think it’s very important and it makes me think of the discussion you had about benchmarking in your podcast a couple of weeks ago, or which was that we suspect the right way to pose the question is to say, “What is the energy consumption for a given desired result for the end user?”

Because there’s a lot of questions of a playoff between quality and energy consumption at every scale inside the quantum computer. What really matters is, can’t the quantum computer solve the problem you want it to solve? That’s your fixed constraint, and then you can optimize everything with under that constraint, but you need it to work. There’s no point having a quantum computer that consumes much less but doesn’t do the job it was built for. That’s clear. So, we take that approach, so we actually called this our ‘metric noise resource approach’. So, the metric is what you want the computer to do as an end user, and what requirements that imposes on the quality of the qubits. The noise is then the modeling of what all of the effects that disturb the qubits from doing their job. Then the resources are the things you bring to stop that noise disturbing the qubits. So, it’s these three principle elements which are a philosophical basis of these more complicated models we’re trying to build to model these systems.

Olivier: What’s interesting is you can add one fact that there’s some economics involved there, because we know that if we want to scale control electronics for example, we will have to have a new economics volume market for quantum computers because that’s what will drive those vendors to scale down the size of the electronics, the power consumption. When you increase the number of qubits, so, it’s about having a new Moore’s for the control electronics driving the scalability of those control computers. It’s a mix of science, technology, and economics.

Yuval: I wanted to ask you, Olivier; you mentioned earlier that there are several hardware vendors. I mean, you mentioned IQM and maybe others that are engaged in QEI. To what extent does a hardware vendor consider energy consumption as a first-order problem? Don’t they first have to get the computers to work and do something useful before they address how much energy they consume?

Olivier: Well, it happens that we hear that a lot from customers themselves. So, we have customers who think that in the short to midterm, one of the justifications to move to quantum computing, particularly in the NISQ regime, is going to be a lower power consumption because HPC consumes a lot. So, if you can replace existing classical HPCs and supercomputing with more efficient energy qubits speaking quantum computers, it makes a lot of sense. But of course, to make that possible, all the hardware vendors have to create better quality qubits. We know that. Whatever the technology, whatever the NISQ qubits FTC approach, they need that. But once they will have done that, they will have to look at this as well.

So, it’s dual approach. All the hardware vendors that we have been talking to, even the large ones from the US, they all are motivated to look at the energetics of the system because we live in a world of finite resource where we can’t escape that. We as a community, I mean the whole quantum computing community, and even technologies community as a whole, we have to have a discourse on that. We can’t say we sell new stuff, whatever the cost, we have to say what is it going to cost including energetics.

Yuval: When I go to a store in the US, and I want to buy a refrigerator, it says, “Well, the energy cost for this refrigerator is so and so many dollars per year.” Do you anticipate that vendors will say, “Well, to run our computer, this is what the energy cost will be.” Do you expect that there’ll be standardization around that?

Olivier: Yes, we want even to create some benchmarks around that. It’ll take some time. Not ready. It’s not ready yet, but we’re still already starting to work with the standardization and standards bodies worldwide, like ITE, to launch a discussion about creating benchmarks around the energetics of those systems. But it’s going to be important to follow the approach that Robert described, the M&R approach because it’s not just the poor consumption per se, it’s the poor consumption compared to what you want to achieve from business standpoint, from the algorithm standpoint.

Rob: I think maybe more generally, I agree that if the primary objective is to make a quantum computer, or more generally quantum technologies because we are also interested in extending these ideas to quantum sensing or quantum communication and so on. But we are not experts in that so that we need other people in the community to come and help us or take the ideas for themselves and go with. But if in quantum computing, if we want to make quantum computer that works, that’s clearly the first goal. So, it might be a little bit too early to think about energetic or more general resource consumption, but I think it’s important to be too early than too late. I think for many technologies in the past, maybe we were a bit too late to think about energy or resource consumption. So, if people think we’re a bit too early, I say good, right? I maybe even agree with them, but I think it’s worth already considering the issue at this stage. Yeah.

Yuval: Earlier, Rob, I think you mentioned the algorithm itself. How much does the power consumption of a system vary between its sitting idle and not doing anything, except that all the electronics are on, as opposed to actually running an algorithm?

Rob: So, that depends how much of the electronics you can turn off when it’s not doing anything. So, typically in the superconducting circuit that I know best, even if the electronics are switched on, you don’t send microwave pulses to actually flip the qubits to change the qubits state to do gate operations except when you’re actually doing the gate operations. So, immediately you have a big difference between the amount of power that you are sending down to the qubits from when the qubits are doing something from when they’re just waiting for the next step in the computation. That is one of the critical elements which connects the hardware design to the algorithm, because you have to protect the qubits against noise, the qubits are typically also very well protected against the microwave signals you want to send.

So, have to send very strong signals through all of the filters that protect them against the noise. So, most of the microwaves you send is dissipated as heat, which you have to evacuate with your cryogenics. So, at the moment when the thing is, when you are actually sending signals, there’s a lot more heat to evacuate than when your qubits are just sitting waiting for the next instruction. So, if you have an algorithm when most of the time qubits are waiting for the next construction it’s quite a different optimization problem from when most of the time, most of the qubits are doing some operation. That’s particularly noticeable in superconducting qubits. In other technologies it may be very different. In some cases you have to control the qubits all the time anyway, just to protect them against basic noise. So, then the difference might not be so extreme, but it’s one of the things we’d like to look at.

Olivier: What we know is the way the compilers are being designed with their optimization and transpiring features, is going to play a key role, and at some point we may have to embed energetics as a way to set up and parameterize those compilers. So, software is going to play a very key role in the energetics of quantum computing, that’s for sure.

Yuval: How much collaboration do you need with the quantum computing vendors to create accurate models, and how much collaboration are you actually getting?

Olivier: Well, at this point in time, we need more collaboration, probably more openness, more data on the full stack approach of all these systems. Your question is right on point. It’s not easy. It’s not easy. We want to build some open science approach. That’s why we organize our workshop in November, and we’ll see how vendors get open, but hopefully there’s also a fundamental research being done in parallel with the vendors, which can bring a lot of stuff as well, Robert?

Rob: Yeah, yeah. So, right now, in our first efforts, we took numbers we could find in the published scientific literature, which mostly comes from… Well, some of it came from Google, but a lot of it also came from university and academic institutions. But of course, if we want to make better, more realistic models of what’s really going on, or if we want to open the codes that are developed by the community so that a vendor can use it to optimize their own setups, then that requires, clearly, some back and forth and everyone learning from everyone else. I think that’s one of the exciting things about this subject is getting to talk to people with very different backgrounds and learning what that’s going on, and learning also what other constraints there are apart from energetic constraints because of course, once you’ve built this global modeling, you can put in other constraints, and probably that’s worthwhile investigating.

Yuval: As we get closer to the end of our conversation today, I wanted to ask you about the workshop in Singapore. So, I assume that everyone in the workshop is going to come in and say, “Yes, energetics are important. We want to control and reduce the energy consumption of quantum computers.” But what presentations or tracks do you expect to hear in the workshop? What topics?

Rob: Okay, so we hope to have a variety of people from a broad variety of subjects scanning from very hardware-based things related to control electronics or cryogenics, to very theoretical things about what is even the meaning of energy in quantum systems. So, we’d like to have a nice selection of people from all of those domains, and that everyone comes with a spirit of being willing to explain their little playground to people with a different background. Then the job of the people present, of all of us together will be to try and develop a common language and perhaps a common set of questions, a common set of problems that we can then move forward with in small groups or in big groups. So, to try and create a dynamic of people who are open and interested in these subjects in general. I think that would be my objective for the workshop, and not set any particular constraints beyond that.

Yuval: Olivier, did you want to add something to that?

Olivier: Well, I can frame the workshop. We’re going to have people from different fields, fundamental quantum devices, people working on quantum hardware. The actual qubits, we’ll have people working on quantum algorithms and software. It’s very important to have all these people work together. Also we want to have people from the HPC world and hybrid computing world because any quantum advantage, whether it’s energetic or computional, has to be always compared to what we do today on actual Classiq systems. So, Classiq information processing is important for that respect and also because Classiq information processing is required to control all distributes, particularly when you implement quantum error correction. So, it’s got interdisciplinary approach, and we want to implement that right away in our first workshop, but it’s the first one. So, we going to learn with flying.

Yuval: A supercomputer today, probably the energy consumption scales with the number of cores, right? If I’ve got 10,000 cores versus 20,000 cores, the 20,000 cores are probably going to be nearly twice as much. How is the scaling in terms of the number of qubits? I mean is that exponential? Is that quadratic, what does your model show there?

Rob: Oh, that’s a good question. So, I would say it’s quite linear, so it’s not so different. But what is perhaps very different from traditional computing is that in traditional computing you often talk about flocks or giga flocks and things like that, and those measures of a quality of a quantum computer have no meaning. So, it’s one thing to say it scales with the number of qubits, but how good are those qubits? What is their fidelity? What is their connectivity? What are all of those other questions that all of those other parameters which also affect the energy consumption? So, if for example, every qubits bit talks to every other qubit, qubits needs to have a connectivity with every other qubit, it could easily be that the scaling is much worse than linear because you need the number of wires then needs to maybe scale the square of the number of qubits or something.

So, then the heat conduction of the wires will scale like the square of the number of qubits. So, it really depends a lot on details beyond just a simple number right now, maybe in the future we will get to a single benchmark of quality and then we can just say, qubits the number of watts for this quality. A bit like people do with the green 500 in traditional supercomputers where they give a number of giga flops per watt, I think they give, or the other way around.

Olivier: Yeah, it improved a lot recently. It went in couple years from 20 giga flocks a watt to 52 recently in the Frontier computing system installed at Oak Ridge, I think in the DOE in the US.

Rob: So, we much too early to be able to give any measure like that. But it’s exactly the questions we should be asking. What are the good measures? So, what is the good measure of quality, and what is a good measure of energy consumption? Then can you make a ratio, which is a useful measure of efficiency or is it a multi-perimeter space in which you can’t do anything like that? I think those are very interesting questions and I don’t think we have really clear answers yet.

Olivier: But, you know, that’s good. Because a good research program is about asking the right questions, not having the answers right away. So, we ask a lot of questions to many people and we ask them to work together to solve those problems.

Yuval: My next to last question is a hypothetical question. So, if you guys could have dinner with one of the quantum greats, dead or alive, who would that be? Olivier, maybe your first, and then Rob.

Olivier: Who? Dinner with whom, you said? I got it. Didn’t get it.

Yuval: One of the quantum greats, one of the famous quantum people.

Olivier: Oh, the famous quantum people. We already have many connections, I would say. We don’t miss those connections. What I would like to have is to have a group of the greats of this world, like Jay Gambetta from IBM, Michelle Devereux, and Alexandre Blais‬ from Sherbrook, and also those guys working on electronics from Keysight, from Joy Instruments, from Sikh, Robert McDermot. I would like to have all these people in the room.‬‬

Rob: So, more like a dinner party than a one-to-one?

Olivier: Yes, exactly. Yeah.

Yuval: I gave you the option of dead people as well, but….

Olivier: Oh, the dead ones? Oh, the dead ones.

Rob: Yeah. Well, I think for sure I would be love to be a fly on the wall on the discussions, the discussions that were happening in the 1920s that led up to that, to quantum mechanics between Heisenberg, and also the discussion between Heisenberg and Schrodinger and who really… To what extent, they were inspired by each other or they figured it all out individually, which I not an expert on the history of that, but I would be delighted to know that.

Yuval: How can people learn more about the QEI and get involved?

Olivier: Simple. We’ve got a website, so you Google quantum Energy Initiative, we’ll bump on our https://quantum-energy-initiative.org/. We point to the QEI workshop, easy to find as well. We just check that you can Google it, although the website is very recent, and you can find also the PRX paper from our friend Alex that was published in June. We’ll send you all the links and we can make good use of it.

Yuval: Very good. So, Olivier, Rob, thank you so much for joining me today.

Rob: It was a pleasure. Thank you very much.

Olivier: Thank you for inviting us.

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.

April 24, 2023