Transcript | Hybrid Classical and Quantum Computing with IonQ

The word hybrid often appears in quantum computing and can mean different things. Learn how blending bits and qubits may help companies achieve real benefits today as we move the slider closer to pure quantum power in the coming months and years ahead in this episode of The Post-Quantum World. I’m your host, Konstantinos Karagiannis. I lead Quantum Computing Services at Protiviti, where we’re helping companies prepare for the benefits and threats of this exploding field. I hope you’ll join each episode as we explore the technology and business impacts of this post-quantum era.

Our guest today is the VP of product development at IonQ, Matthew Keesan. Welcome to the show.

Thanks for having me — great to be here.

Yes, it was great — we got to chat a little bit at the Q2B show. That was fun, and Matt’s a fellow New York City denizen.

That’s right. Yes, New York, which should be the home of quantum computing someday.

Yes, right. You feel that it’s, like, we’ve got IBM to the north, and the financial district is here, and —

The demand is there.

Yes. Everyone’s edging in, but — so we’re the outcast here. Everyone else at the show is not living in our area.

The one downside is, the best place for a quantum computer is deep underground, under a mountain or far from cosmic rays or the sounds of regular people.

We’ll just put them in the subways on the weekends, because they’re not being used by trains. They’re always down. Tell us how you ended up in quantum. You were one of the initial — what is it — four employees of IonQ?

No, probably somewhere around No. 12, maybe. It’s a crazy coincidence: When I was in high school, in physics class, at the end of the year, we all had to do a poster presentation on a new development in physics. We all had to learn traditional mechanics, stuff like that, and it just so happened that the year before, Chris Monroe had run the first quantum logic gate ever at NIST in the ’90s, and I was, like, “This is amazing!” I did my whole presentation on quantum computing — got to college. Of course, as an undergrad at that time, you couldn’t study it — there was one class in our CS department — but I followed it for years after, just because it seemed so cool, and then, in a wild coincidence, now I work for Chris.

IonQ was spun out of his research at UMD, and his longtime collaborator Jungsang Kim’s research at Duke, and it happened to be funded by one of the early investors at my first startup. I’m a software engineer. I’ve been doing startups for the last 15 years or so. It came in through a coincidence, and when I heard it was Chris, I was, like, “This is amazing. I’ve known about you for 30 years at this point.”

Yes, same here. I was super excited to have Chris on the show a few months ago, because, yes, I’ve known about him forever. It was like having one of the founders of the whole industry.

Yes, and his strong Midwestern-dad vibes. You’re, like, “You are that elder statesman of the whole field at this point.”

Yes. The reason I wanted to have you on is, just recently, you guys announced that you’re going to have a platform with Dell delivering hybrid classical quantum. I wanted to talk about the whole meaning of hybrid in general, what it means to use the two technologies together, and then we can get into your approach.

Broadly speaking, today’s quantum computers, as we know, are noisy, and the best way to get value out of them is to pair them with a classical computer that still does some things well by keeping the things that a classical computer does well on the classical computer and leaving the quantum stuff to the quantum computer — get more value out of both, if that makes sense. Just as we use GPUs today to render video and 3D images for video games and so forth, we offload that from our CPUs. By offloading certain parts of certain algorithms onto our QPU, we could start to see a path toward quantum advantage.

That’s a good point, and going toward quantum advantage, most companies already are hearing that they don’t have it now. It’s not something that they can get today, and everyone’s hoping for that great moment. Do you feel that having the cyber approach will ease that self-imposed barrier? Some companies are, like, “There’s no quantum advantage? I don’t want to touch quantum.”

Yes, and to those companies, I would say, it’s equivalent to being in 1975 and saying, “There’s no computer advantage — I’m just going to keep doing my accounting on pieces of paper,” and then suddenly, VisiCalc comes along, and the companies that are ready to adopt personal computation in the late ’70s suddenly have a real advantage. To answer your question explicitly, yes, the innovation of hybrid classical quantum computing was specifically designed to answer the question “How can we get value out of quantum computers today before we get to scaled, fault-tolerant quantum computing?”

One way to do that is to combine what your quantum computer can do with your classical computer. It eases the transition in a very literal sense of getting more value out of quantum computers sooner before we hit fault tolerance, because I am the least part of the collection of people who believe that we don’t have to have fully fault-tolerant quantum computing to get business value. Second, it means you can take your classical computer, which you already need for all sorts of other stuff, and pair it with today’s quantum computers and start to pave the way for those scaled-up applications of the near future.

What I’ve found interesting with this approach is, you map it out. You show in block diagram all the different areas, like runtime, runtime execution, the ability to hand off the simulators when it actually goes to, let’s say, Aria for a true QPU. A big initial question is, there is transparency here, and how would you feel this compares to something like, let’s say, D-Wave’s hybrid solver? We’ve used it. I wrote a paper using it, and I still don’t know how much of it was quantum. It’s like a black box. You throw something in, something comes out.

Yes, that’s like — are we allowed to make this joke? I feel like the age-old question is, how much of that is really quantum?

Yes, I’d love to know.

I’m certainly not qualified to answer that question. For us, we want to be totally transparent that, yes, the quantum computers of today are simulatable using commercially available GPUs at best, and leaving this idea of some of these very specific, not generally useful experiments that you might need a more powerful computer to do. They’re certainly simulatable by classical computing available on Earth. We can say that there’s no quantum computer we know of on the planet that can’t be simulated right now, but to develop algorithms, you need to validate them on real computers.

Ultimately, if you believe, as I do, that there will be quantum advantage before we have fault tolerance, that means we need to build algorithms and error-mitigation techniques that are resilient to real noise on real machines. That will likely be a hybrid approach combining classical and quantum, and that’s why we think there’s value to customers today, because you could start using, in these smaller systems, the same algorithms you’ll be needing to use at larger scale.

To help people visualise how they access something like this, is this available on IonQ’s quantum cloud?

Yes. Through the partnership with Dell, they are providing classical control software, etc. Then all you need to integrate is an API key with our direct access.

Is any of this physically colocated? Did Dell have to do anything like move hardware closer to, let’s say, Aria or anything like that, or is there no benefit to having any kind of colocation?

We are definitely excited about working with partners interested in colocation, but to develop the actual integrations in the software, it doesn’t matter. That’s a variable you can tune. If you buy a quantum computer and a classical supercomputing platform, then you put them close together, there’s clearly a real improvement purely from a speed-of light standpoint. However, to actually build the platform, we didn’t have to physically do that. We do have Dell hardware nearby, but just because Dell makes a lot of computers and they’re great.

Yes, that’s true. It just senses, “There’s Dell nearby. This’ll work better now.”

Exactly, but, yes, in all seriousness, there’s no technical reason it has to be close from a purely functional standpoint, but from a user-experience standpoint, we certainly get much faster iteration time the closer those are together.

There are simulators running also as an option as part of this whole workflow?

Yes, and it speaks to your experience with D-Wave. We make simulation available from a software standpoint for free because we want you to be able to preflight your programmes, be able to test them out, feel like they’re going to work from a purely logical standpoint before you actually run them on real hardware.

And some of that will run on Dell servers, and then some will run on machines you made?

Our simulator, at small scales, it doesn’t matter whether it’s on a CPU or a GPU. At larger scales, you’d want to use GPUs that aren’t made by Dell but are packaged into the Dell hardware. Inside your Dell server, if you have, let’s say, a k80, conventional, reasonably affordable GPU, you can easily simulate up to 29 qubits.

That’s great. Is this all intelligently handled, or do you have to make any decisions on your own? Does it decide what the flow will be for you in the process — how much hardware will be needed for simulation stages, when it would make sense to go to Aria? Is there anything that’s given to someone who’s trying to run a problem to guide them through that, or do you make all the decisions manually?

Dell’s tooling is designed to help with that, but of course, we also expect that researchers perhaps like yourself also want to have a lot of control and make some of those decisions themselves. It depends on the sophistication of the customer, what’s most appropriate there — whether taking fine-grained, low-level control or working at a higher level and letting the decisions be made automatically.

Has there been initial benchmarking done, or anything like that, to show the benefits of this approach with the runtime?

We haven’t done any formal benchmarks yet, but it’s definitely something we’re excited about for the coming year.

I’d love to know more about when people use IBM Quantum Cloud to access your machines directly. Do you want to talk about those benefits? You could, obviously, get access to IonQ on all three of the big cloud platforms.

Philosophically, all of our partnerships speak to this. We’re excited about democratising access to quantum computers and meeting developers and researchers wherever they are, because, as you know — as your podcast discusses — we’re in the very beginning of the whole field. We have all sorts of exciting results and reasons to be excited, but it might as well be 1945 in regular computing.

There are so many things we take for granted today that would’ve seemed like magic to the early makers of computers. From our perspective, the more places we can meet people and give them access to our hardware, the better, because that’s where the future innovations will come. To your specific question on, why work with us directly, one reason is something you asked about at the very beginning, which is transparency — you’re very close to exactly what’s happening.

In terms of benefits on the IonQ quantum cloud, one of the things I’m most proud of is, we believe we have perhaps the best, or among the best, in the world, optimising compilers for trapped-ion hardware. Although you can work with many other compilers through our native gate interface, bypass our compiler stack and play with your own compilation or play with compilation provided by Qiskits or through others, the native compilation we do in our testing has outperformed all other comers. That optimisation is huge, because that means getting the most benefits from today’s noisy, intermediate-skilled quantum computers. Another one is being able to work directly with us on support issues without any layer in between. If you need help optimising your programme or making it better, we’re there to answer questions and improve the likelihood of success of your research.

You also get access to a machine that’s not available anywhere else. You’ll be able to hit Forte this way?

Yes, long-term, of course, we always want to make our computers as widely available as possible, simply because it behooves us, and it’s to everyone’s benefit for us to do private betas before we make them publicly available.

Yes, even though we were kidding around before, in many ways, is this similar to the D-Wave approach? You can get them on another cloud, but if you wanted the hybrid solver, you have to go directly through D-Wave for that? And then it would make sense that for the hybrid here, you would go directly, and for the latest machine, you would go directly. Tell me more about the optimisation that happens. Is it different? When we talk about optimisation that occurs, let’s say, in the hybrid approach, is this a completely different hybrid than hybrid optimisation that’s going on here in the cloud, to get circuits to run better, etc.? Or is it the same thing?

From a compilation perspective, the circuits are the circuits, and whether they’re running in a hybrid context or not, they get the same best-in-class optimisation. Where the hybrid mode has benefits is in scheduling and throughput. If we know that what we’re running is a hybrid workload, we can optimise the flow through the system to be simply faster. It’s mutually beneficial, because it’s a better-utilised system for us. That means we can do more workloads overall, and researchers get their answers more quickly. In the context of, for example, the public clouds today, that’s to protect us from the data that’s opaque to us. We can do certain optimisations that we’re able to do when we know, here you come with your, let’s say, 1,000 iterations of some hybrid algorithm you want to run.

How would you compare this to, let’s say, the optimisation that goes on with IBM? For example, IBM does do a little bit of that last pass on your code before it goes to the machine to make sure you’re not doing silly things like wasting gates and wasting qubits and repeating processes. It optimises it lightly. But then there’s that whole idea of going down to the pulse level on a transmon machine to code those. Is this like a replacement for that, like having any special ET knowledge, any expertise, and going down and controlling the optical tweezing going on, or is there any benefit to going deeper than this optimiser, or is it doing all that good stuff for you?

Yes, we continuously improve how we do our gates. Our pulse solutions, as we call them — the shape of those pulses that perform our quantum logic gates — those are undergoing continuous improvement, and that’s an orthogonal process to optimising what sequence of logical gates is going to realise the programme. We’ve invested in more optimisations than IBM. IBM’s approach, as I understand it, has historically been, “We’ll give you access, and you do the optimisation you want to do,” which is great, and we support that for customers who want to plug in at that level. They’re welcome to use our native gate interface, bypass our compiler and run on their own compilation, but for most customers who are not researching compilers, but just want the best possible performance, we have a variety of techniques that we deploy simultaneously to improve compilation as well as perform error mitigation at the end of your circuit.

It’s possible that you could run the same code six months apart on the same machine and get better results.

Yes, and in fact, that’s our goal, and we’ve shown that this year. It’s not yet available to the public, but as you may know, the way we benchmark our computers is with a basket of algorithms that produces what we call AQ, or algorithm qubits. Benchmarking quantum computers is a whole subject in itself and there are a lot of different approaches. We believe that what matters is, can I actually use the qubits to run a useful algorithm?

We decide all these vanity metrics about very specialised algorithms or whatever, take a basket of real-world algorithms, plant them together and say, “What are the most qubits that we can actually get answers out of that are useful?” And that means that our circuits have to be quite deep to run all the algorithms in this basket of algorithms called AQ — and you can go on our website or Google to read about the technical details. But over the course of this past year, our top-performing system has moved from 21 algorithm qubits to 25. Every single increase in AQ, that’s a doubling of performance, and all of that has come from making these kinds of tweaks improving gate solutions, tweaking the control routines, how we calibrate and so forth, improving compilation. All that stuff improves our ability to run real-world algorithms.

Yes, that’s definitely proof. It’s the same little ions — they are just being used more efficiently.

Yes, and generally, generation to generation, obviously, you’ll see the biggest improvements, but in the process of bringing a particular system to the public, we want to be continuously improving for as long as possible.

Is this the kind of thing where if someone were to buy one of these machines to have at their own location, would they be able to constantly take advantage of the same approach? Would it be some software update or hardware update?

Yes is the simple answer. Like with any production system, obviously, we don’t want to just be updating software or hardware on your premises without your permission, but certainly, just like with any product, there’s a stream of updates that you’ll take until the system becomes end-of-life, and that could easily include software as well as hardware.

Is there anything you can share about Forte and how it’s going, its performance targets, and what the road map on that might look like?

All I can say is, bringing Forte up and getting it into early access is going well, and it’s super exciting. Forte represents the first of a new generation of control, as we’ve talked about earlier this year, for us, and so getting it into the hands of customers is exciting.

I’m looking forward to playing around with that.

Hopefully, you’re on the shortlist.

Yes, that would be great.

If not, I can pull some strings.

Well, I would hope so. That’s awesome. I want to use that on some coming experimentation, for sure.

What does your gut tell you would be the first application to show advantage with a pure like QPU?

I will tell you my good answer, but also, my disclaimer is that there have been demonstrations over the past couple of years of very contrived one-offs where you could say there’s an advantage, and it’s an advantage that you never wanted. If we put aside the contrived algorithms, I’m leaning toward machine learning or maybe chemistry.

I don’t know if you’ve seen this: We have been training image-recognition models in Aria where we’ve loaded a little picture — not a lot of qubits comparatively, but a recognisable picture, in terms of pixels — into atoms and used that to train machine learning models, and if you query the atomic state, you can see these pictures that have been stored in atoms. It’s completely mind-blowing. There’s no advantage in that yet, but that field is progressing. Quantum machine learning seems to be moving extremely quickly right now, and it wouldn’t shock me for approval advantage to come there first.

Yes, that’s been my gut for gate-based quantum computing. I thought it would be something in quantum machine learning. That’s what we’re going to see first, too.

What’s cool about that is, that’s such a generalisable technique in so many areas, and as soon as the first domino falls, there are a lot of people waiting to cross that chasm who won’t have a hard time translating their existing workloads into quantum libraries and so forth.

Yes, I definitely agree. Do you foresee, on the IonQ quantum cloud, having reusable code for people who are getting started?

Right now, we provide a lot of notebooks and examples, but they’re certainly not libraries, let’s say. In the long term, I would love to see that. We’re a small company and a small team, but we do think, in the long term, we can optimise the algorithms for our hardware, at least today, better than anyone else can. It makes sense. We co-design software and hardware together, and the further we can go up the stacks, the more value we bring to customers.

This is definitely something I’m excited about, particularly because as a software engineer, as much as I genuinely love assembly programming, what makes a software engineer powerful is leverage: You don’t have to think about ones and zeros when you’re programming. You can make 3D video games using libraries that completely abstract away GPUs and CPUs — so many aspects of what goes into actually rendering a three-dimensional object on your laptop or whatever.

When that exists for quantum computing, suddenly, as we’ve been discussing, all sorts of avenues for quantum advantage become accessible to everyone without having to know anything about what’s going on inside the quantum computer. And that’s when quantum computing, to me, has achieved its promise — not when today’s researchers can do things that have never been done, as cool as that is, but when ordinary people can reap the benefits without having to know anything about what’s going on inside the box.

Yes, that’s a great point. Yes, a great place to end — that idea that you’re writing the code for your machines, so it would be great to have those libraries widely available to push this forward.

Yes. All of us, that’s what we’re excited about — quantum computing changing the world. We believe — and I certainly believe — IonQ has the world’s best quantum computer, that we know of anyway, but the bigger picture is even more important, which is what this technology could do for the whole human race.

Yes, and with that, I’ll thank you again for joining. This has been great — I enjoyed it.

My pleasure.

Now, it’s time for Coherence, the quantum executive summary, where I take a moment to highlight some of the business impacts we discussed today in case things got too nerdy at times. Let’s recap.

In many ways, all computers are hybrids of different types of processors. In your laptop, for example, there’s a CPU and a GPU of some sort combined with other controllers and so on. They work together to route jobs where they need to go. Hybrid classical quantum computing is very similar: Give to classical what belongs to classical bits, and give to quantum what it can handle best with qubits. D-Wave uses this type of approach with annealing in its hybrid solver.

IonQ is doing something similar in the gate-based world. IonQ’s most powerful quantum computer currently available to the public is Aria. It features 25 algorithmic qubits. IonQ has partnered with Dell to combine the power of Aria with PowerEdge servers. The approach will allow users to test code on powerful simulators and run algorithms on Aria. The environment emphasises transparency of how the code is running and is accessed through IonQ quantum cloud, which brings advantages like circuit optimisation. I’m a fan of hybrid approaches, because I feel that companies looking to get into quantum have the best chance of seeing benefits in NISQ era.

That does it for this episode. Thanks to Matthew Keesan for joining to discuss hybrid classical and quantum from IonQ, and thank you for listening. If you enjoyed the show, please subscribe to Protiviti’s The Post-Quantum World and leave a review to help others find us. Be sure to follow me on Twitter and Instagram @KonstantHacker. You’ll find links there to what we’re doing in Quantum Computing Services at Protiviti. You can also DM me questions or suggestions for what you’d like to hear on the show. For more information on our quantum services, check out Protiviti.com or follow ProtivitiTech on Twitter and LinkedIn.

This is our last episode of 2022. I hope everyone has a great holiday season, and we’ll see you in 2023. Until next time, be kind, and stay quantum curious.