Podcast with Oliver Dial, Quantum CTO at IBM

Transcript

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

Oliver Dial:
Hi, it’s a joy to be here.

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

Oliver:
My name is Oliver Dial. I’m a physicist with IBM Quantum. We’re a team at IBM that’s trying to build quantum computers and make it real for the world, useful for the world.

My job title this week is Chief Technology Officer for IBM Quantum. But really, on a day-to-day basis, I’m worried about making sure that our quantum hardware and our quantum software work well together and deliver a delightful experience for our users.

Yuval:
Congratulations on the CTO side. Does that mean that you cover all aspects of IBM’s quantum activity on the technical side or, as you said, focus primarily on the hardware?

Oliver:
Well, I don’t think anyone can really cover all aspects of this. At the end of the day, bringing useful quantum computation to the world doesn’t just mean having the best and the biggest quantum processors, but it also means having the software that means you can use it well, as well as advanced techniques like error correction and error mitigation that let people get kind of the right answers out of it.

And so it really takes an entire team to deliver these technologies to the world. So just like anyone else, there are places where I’m better and places where I’m worse. And I’m most comfortable in the hardware, but I look at it all.

Yuval:
By now, IBM has several or many quantum machines with a different number of qubits, and you keep introducing new chips once a year. I mean, wonderful progress. If you were speaking with a customer today, do you always point them to the highest-end machine, or do you say “well. this machine has better error performance or this one has different connectivity or other features?”

How would you help customers select which machine they want to target?

Oliver:
That is a wonderful question. We really think there are three things that we need to improve about our machines to bring this useful quantum community of the world, and we talk a lot about scale, quality, and speed. Scale is just the number of qubits. If you don’t have more than about 50 qubits, you’re not even beyond what a classical computer can simulate, so you’re sort of not even in the right ballpark for working in quantum computation. Quality, the error rate that those qubits have, the quantum computers, the way we use them today are analog, and so every time you do a computation there’s some probability that you get the wrong answer. Higher quality means you get the right answer with higher probability. And then speed, just how quickly can that processor give you the answer. We work on these three problems in parallel.

We have projects aimed at delivering the biggest processors and of course, those get the biggest splash in the headlines, but we also have projects aimed at improving the quality and the speed of our quantum devices. And so at any one moment, our biggest devices tend not to be our best in terms of quality. Our best in terms of quality tend not to be the fastest and our fastest are almost never our biggest. And so kind of depending on what the client’s interested in, they may be more interested in the biggest device, or the best device, or the fastest device.

And really, we’re also extremely interested in technologies like error mitigation and error correction that let us trade these against each other, running more circuits to get a more accurate answer, for example, using error mitigation.

But at the end of the day, you have to remember these devices are not yet useful. They don’t solve our clients’ problems at the scale that they’re interested in yet.

And so the biggest thing that we do with all this is we point at our roadmap, which is sort of looking into the future as to how we’re going to continue to improve scale, quality and speed for our customers.

And so when we’re talking to our customers, the conversation is not just about the devices that we have today, but also about building the confidence with them that we’re going in the right direction for the devices that they want tomorrow.

Yuval:
One of the devices that you have, I think, is it 27 qubits, maybe called Falcon. Do you go back to that machine and say, oh, now we’ve got better error mitigation or different pulse control, and does it get better over time or is it just, here’s what it is, and now we’re focused on the newer models?

Oliver:
They absolutely get better over time. They get better in two senses. One is, you mentioned error mitigation, pulse control as our technologies for using the devices get better. The quality of our users’ experiences with them gets better. One of the recent things we’ve done on our larger Eagle devices is actually changed the way that we calibrate them.

Previously, we tried to calibrate them to provide the absolute best possible performance on a small subset of the qubits. And that was really good if you’re working on things like quantum volume, where you wanted to use a couple of qubits and run a really complicated circuit. And we’ve changed so that instead, we’re optimizing for the average performance of the device as a whole, which is much better for techniques like error mitigation, where you really want to use the whole device, and you’re limited by the worst qubits.

The other sense in which the devices continue to improve is we spin new hardware revisions. If you go look at the IBM Quantum web page, you’ll see next to each Falcon processor, there’s also a revision number. And those revisions go all the way up to R9, actually. So we’ve done nine different versions of that Falcon processor with continuous improvements in connectivity, in coherence time of the qubits, and in error rates.

Yuval:
If I’m a commercial customer using IBM beyond what you guys allow access for free, I know you have multiple machines, but for a given machine, how many hours a day is it available on average? Does it need to be calibrated every two hours and taken offline? Does it happen once a week? How does availability look for a particular single machine?

Oliver:
So right now, the number sits at around, I think it’s 15% of the time per day on an average-sized device gets spent in calibration. It’s a little bit larger for large devices. It’s a little bit shorter for the small ones. A large fraction of that time is actually spent benchmarking the device.

If you go to the Quantum Experience web page, there’s error rates for all the gates and things like that. So a substantial fraction of the time is just updating those error rates once a day. And there’s also smaller things like calibrating small amounts of control and readout drift in the device that happens a little bit more frequently.

From the user’s viewpoint, this all happens transparently. The calibration jobs are interleaved with the user jobs. And so it’s not like the device goes offline for two hours a day during calibration. But this is kind of a continuous process throughout the day.

Yuval:
Beyond the IBM Cloud, I think IBM sold several machines. I think there was a press release about the Cleveland Clinic and maybe the Fraunhofer Institute. and I think once upon a time, a machine was sold to Japan. Why would a customer buy a machine? I mean, doesn’t it get obsolete in a couple of years?

Oliver:
So first of all, I think it would be more correct to say that they’re leasing those machines as opposed to buying them. That for a variety of reasons, we don’t sell these machines to anybody right now. It just creates a lot of complications. Among other things, as you said, why would you actually want to buy one anyway?

There’s a lot of reasons why people choose to lease these machines. Some of it is they want just exclusive access to a machine, whereas if you’re one of our premium customers, you’re of course sharing access to a set of machines with a set of places.

Part of it is, a lot of the reason people are using our systems today is to get what we call quantum ready. Imagine somebody told you five years before it happened that IBM System 360 was coming, and if you could get a team of programmers ready, that knew how to use that, then on day one, when it came out, you would be able to take off compared to your competitors.

That’s sort of where we see where we are today in quantum. And so, from our clients’ viewpoints, it’s not enough just to have the machine, but they have to have the interest and the excitement of the organization around using their machine. And there is nothing that builds that interest and excitement quite the same way as actually having the physical thing in your building, where if you’re a university, your students, if you’re a company or a public organization or a national lab, your employees can see it, can look at it and kind of be inspired by it.

Yuval:
And speaking of the human aspect, when you lease the machine, do you also “ship” a couple of technicians with it? Or is calibration remote? Or how does it work operationally?

Oliver:
You know, the first one of these lease machines that we stood up was during the middle of the COVID pandemic. And so we had people on their laptops, on WebEx, communicating remotely with the researchers at, in this case, it would have been the machine in Germany, walking them through the steps to set up the cryostat and to hook up all the wiring and hook up all the hoses, and then a couple of IBM employees to install the actual device.

So we’ve done this all sorts of ways. And it’s kind of huge kudos to the team for managing to get this done under difficult circumstances. But yeah, typically, there will be IBM employees to handle the actual details of the final setup, the calibration, all that kind of thing.

Yuval:
Do you feel that most users still have to be PhDs in terms of their education, or beyond IBM Quantum Experience, do you find that actual commercial users can broaden the types and education of people who benefit from the machine?

Oliver:
There are different types of users. If you really want to work on the low-level technologies like error mitigation and error correction, then I think really having a deep understanding of the type of, how these computing systems work, and the types of errors that they have and how noise influences them is essential. So for that sort of kernel-level developer,

I think you still need to be not necessarily a Ph.D., but at least very deeply enmeshed in quantum information science and the physics of these devices.

But at the other end of the stack, as an application developer, we have pretty well-developed libraries at these points to run particular quantum algorithms and particular problem domains. And at that level, what our users are really bringing to us isn’t the expertise of quantum computer, but the expertise in the problem domain that they want to work in, whether it’s chemistry or high energy physics or finance.

And so there, being a Ph.D. in physics is just utterly irrelevant because that’s not the knowledge that they’re bringing into this collaboration. So the knowledge that they’re bringing into this collaboration is their own technical background in their field.

Yuval:
I want to talk a little bit about quantum computers at the HPC center. So I want to start with energy consumption, actually. One of the selling points or the advertised selling points of quantum computers is they can do calculations that HPCs do today for a fraction of the energy. But when you look at a system like yours with cryogenic cooling and so on, how much are you worried about the actual energy consumption of the quantum computer? And are there efforts to try and reduce the energy footprint of the machine?

Oliver:
That’s a wonderful question. I’m gonna start with a sort of useless answer for you. That for some problem domains where it’s basically possible to solve the problem with a quantum computer, but a classical computer, no matter what the scale you chose to make it, would never be able to solve it. The sort of energy efficiency becomes kind of a divide by zero problem, right?

It would take an infinite amount of energy to solve this with a classical computer. So, of course, the quantum computer is more energy efficient. And I think the first places where we find advantage are going to be more in that domain, where because the classical computer that you would need to solve the problem is beyond anything that we can build, the amount of energy that the quantum computer takes as long as it’s sort of human-scaled is almost irrelevant.

That having been said, we’re very worried about power consumption as we try to continue to increase the scale of our quantum computer to reach this era of useful quantum computer with error mitigation and error correction. Actually, a really nice example is, you mentioned refrigeration.

I’m gonna switch the topic a little bit to control electronics, because that actually turns out to be one of our big wall power drivers on our systems today. When we first started deploying IBM Quantum Systems, I think it was seven years ago with this little five-qubit device in the IBM Quantum experience, we used just commercial off-the-shelf control electronics.

Our best estimate is it took us about 300 watts of power to control each qubit. A couple of years later, when we were making our 20-qubit penguin devices, we’d advanced that technology a little bit. We were still using off-the-shelf electronics, but we’d gotten down to about 70 watts of electronics– 70 watts per qubit. In our latest generation of electronics, our so-called Gen 3 electronics. We’ve dropped that all the way down to 35 watts per qubit. And so we’re making actually pretty substantial strides in how much power that it takes to control our devices.

But when we look into the future, 35 watts per qubit is still pretty daunting if you want to talk about a 100,000 qubit machine. Not undoable, but a little bit scary. But we’re actually investing in what we call cryogenic CMOS controllers. So these are chips that are responsible for controlling the qubits that actually live inside of the cryostat at a higher temperature stage. And because they are closer to the device, where they can use much lower power levels to communicate with it.

And honestly, because they’re based on ASICs, application-specific ICs, as opposed to FPGAs, which are sort of all-purpose devices, we’re getting that number down to the 10 milliwatts per qubit range with these CryoCMOS ASICs. Now, this is a research project. This isn’t something that we can get out to our customers today. But it’s something that we see sort of as being the future of the technology.

Yuval:
Some companies believe that quantum computers should be application specific, that there’s a different architecture or connectivity that’s better for machine learning as opposed to optimization or something else. Do you subscribe to that notion or do you think that we just have to make one type of quantum computer and that’s going to be pretty much best for all applications?

Oliver:
So there are two types of application-specific computers people talk about. You mentioned one where maybe you designed the connectivity different depending on the algorithm. People also talk about sort of quantum emulators where you build a specific chip to emulate a specific physical system. If we just focus on that first part, the quantum computers, in some ways, look a lot like classical computers. That as long as you have sufficient connectivity, that you can use one connectivity of your chip to emulate a different one with sort of polynomial overhead. And so it means that this isn’t really a cut-and-dried trade space where you just pick whatever topology you want and you’re done. That it’s a trade-off between what topology can you get that best performance out of and what topology does your application need and how expensive is it to map between them.

And a really great example of that, actually is I mentioned our 20-qubit Penguin processors from a few years ago. Those had a square lattice connectivity. So each qubit was connected to four neighbors. It turns out that in the technology that we use, that produced the yield problem, that we had a lot of on-chip, what are called frequency collisions.

And so a lot of the qubits honestly didn’t work that well. We realized that we could actually reduce the connectivity very slightly to this, what we call heavy hex pattern, which is a mixture of connections with three qubits connected to each other and two.

So there’s a lot less connectivity, but the error rates and the yield got enormously better. And so that’s sort of showing that by changing the topology of the device, we can improve the performance of it.

The flip side of it, the motivation for that square lattice layout was an error-correcting code called the surface code. And we were able to develop a new error-correcting code on this heavy hex lattice that would be able to take advantage of this new topology.

So the short answer is, I think it’s a trade space question, whether your general purpose processor with a layout that works best for your technology is the best choice for your application or not. But we’ve generally found that we’re able to co-design the application and the processors to provide the best possible performance.

Yuval:
Beyond a certain scale, word is that QPUs have to be interconnected. So you’re not going to have one monolithic QPU that is going to have 100,000 qubits. So what do you see as the trade-off? So, for instance, am I better off using a 400-qubit chip, or am I better off interconnecting 400-qubit chips?

Oliver:
I really hope you don’t care. that some of the big drivers toward modular systems are around manufacturing. That ther are sizes of chips that we can easily make using semiconductor technologies. And so that kind of puts an upper limit on the geometric size of our chip. There’s yield associated with each component on our devices.

And so that says as we put more and more qubits on the chip, the manufacturing gets less reliable. And there’s also an awful lot of IO required for our devices. that also eats into this chip footprint very quickly. And so at any given stage of development of this technology, and I hope the number gets bigger and bigger, but there is a specific number of qubits that we can put on a chip that is going to maximize the number of qubits that we can produce per wafer. It’s gonna be sort of the best choice for the size of the chip. And so as that size of the chip marches on and on, we’re gonna be putting more and more qubits onto this module, but it’s not necessarily going to match what you as user want.

And that’s where these modular technologies that allow us to join multiple chips together really come into play, because they let us eat up the mismatch between the size of the module that I can most efficiently manufacture and the number of qubits that you as a user want in an effective and efficient and high fidelity way.

So if we’re doing this correctly, it’s not that you want a specific number of qubits on your chip, it’s that you want a specific number of qubits. And then it’s up to us to find how to use our manufacturing process to deliver that device in the best way. Now, if you go back to our roadmap, for the last couple of years, we’ve really been focusing on that first problem.

How many qubits can we cram onto a chip? Two years ago, we released Eagle with 127 qubits. Last year, we released Osprey with 433 qubits. This year, we’re coming out with Condor with 1,000 qubits per chip. And that’s really us exploring the upper end of the qubits per chip scale.

Next year, we’re really going to begin exploring this modular aspect with the introduction of chip-to-chip couplers. Think of it as a multi-chip module if you’re used to classical processors, and also long-range module-to-module couplers. As a user, I hope that you’d never notice the chip-to-chip couplers, that it just looks like one big processor to you and you don’t care. That may be a little bit aspirational for the first releases, but we’ll get there.

Yuval:
As we get closer to the end of our conversation, just a couple more questions. When you look at classical CPUs, initially, they had numeric co-processors, right? The math co-processor was ultimately integrated into the main CPU. When you think about quantum classical integration, do you ever see a situation where the same physical computer hosts both a classical processor as well as a quantum processor, or will they forever be two separate machines, maybe in close proximity to each other with some high-speed interconnect?

Oliver:
Yeah, so this is actually an area of research that we’re really interested in the next few years. and we’re talking about this as the transition from quantum computers to what we call quantum-centric supercomputers, which is the recognition that if we’re really looking at problems of a scale where quantum computation becomes important, that we’re not looking at purely quantum workloads, but we’re looking at mixed classical and quantum world loads, and we need to be able to make these two worlds work together, which is really challenging because the speeds and the data rates are just massively different between quantum computers and classical computers.

And so, how you connect these two things together is a little bit unclear at this stage. That having been said, you mentioned that your computer has a classical coprocessor. But what you might not be thinking about is you use cloud compute a lot more than you realize. For example, when you pick up your phone and say, hey, a particular name that I’m not going to say right now because my phone would respond to me, your phone’s not doing that voice recognition. That’s going off to a big classical compute resource in the cloud.

In the same way, I think for a long time to come when people want to consume quantum computation as a product, they’re going to be using it as a cloud resource, just because it’s more convenient and more cost effective to share these big extensive machines between many users than it would be to try to put one into your home. And fortunately, the cloud model works very, very nicely for quantum computers because as I said before the data rates and the latencies associated with them are at least today much longer than classical computers anyway.

Yuval:
On a professional level what keeps you up at night?

Oliver:
Well, first of all, let me say I am deliriously happy right now. You may not be able to tell from my voice I’m deliriously happy right now. And the reason is just the progress that our team has made, and the field has made in the the last couple of years is just amazing.

As far as scaling the processors go, I am unbelievably stoked about…, we’re coming out with a processor called Heron this year, which has enormously higher fidelities than anything that we’ve produced before. So enormously higher quality.

I’m extremely excited about error mitigation because I think it brings the era of useful quantum computation far closer than we ever thought it would ever be.

And it’s a little bit technical. I’m interested in some new classes of error-correcting codes that I think will bring fault-tolerant quantum computation far closer than we ever thought it would be. Probably the thing that worries me the most is making sure that everyone’s expectations remain realistic. That you can find an awful lot of people that are really excited about this technology, which is wonderful.

You know one of the things I do in my spare time is I teach some wave classes to elementary and middle schools students and you know It’s always it’s a little bit challenging sometimes to get people excited about waves.

But there is no challenge to get people excited about quantum computers. You say those words that people are just like, oh that sounds cool, I want to do that. I want to know about that. But the reality is that we are still not where we need to be as far as solving useful problems for clients. And to me, a client is the entire world.

Don’t think of that as a business thing. Think of that as anyone outside of my team using a quantum computer. And I’m a little bit worried about the lag between sort of when people expect that based off of popular media, based off of the news, and when it’s really going to get out of the lab into the world.

Fortunately for me, the best way I can fight that is by trying to make sure our devices get better as quickly as possible. And so that really aligns well with our team’s goals. But it does worry me.

Yuval:
And the last question, a hypothetical, if you could have dinner with one of the quantum greats, dead or alive, who would that be?

Oliver:
Oh my it’d have to be Feynman. This may sound dumb, but it just comes back to the human aspect that of all the people that I think of when I think of big advances in quantum mechanics, I think he is the person that has the personality that I would most be interested in having dinner with.

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

Oliver:
Thank you so much for your time, it’s been a wonderful conversation.

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.

May 22, 2023