In a rather remarkable way, and despite the frequent hype, the behind-the-scenes work of developing quantum computing has dramatically accelerated in the past few years. DOE, NSF, academia and industry are all in the race. One start-up – IonQ – recently went public via the SPAC route currently in vogue. The Quantum Economic Development Consortium (QEDC), now run by SRI international, is, among other things, pursuing standards development. The National Quantum Initiative Act has spurred creation of a number of research centers. DoE and DoD have even waded into the quantum ‘chip’ foundary business.

Make no mistake. Classical systems are still faster and more accurate than quantum computers for most of the problems being targeted, but not by so much and maybe not for much longer given the effort being poured into quantum research. Quantum chemistry is a frequently-cited poster child application for QC. Consider this observation from Raphael Pooser, a PI for DoE’s Quantum Testbed Pathfinder project and a member of Oak Ridge National Laboratory’s Quantum Information Science group.

“Two or three years ago, we were seeing that we could work really hard to get interesting results on quantum chemistry out of the quantum computers of the day. The concept of chemical accuracy, which is sort of the gold standard, was in a nutshell very hard to attain on the hardware if you didn’t have an in-house device that you’d built yourself. Fast forward to today, we just got through running this benchmark on the latest quantum computers from IBM, and we have some unpublished results from other devices. These systems’ performances have grown by leaps and bounds. It’s gone from being very hard to achieve chemical accuracy on those same problems three years ago to becoming routine now,” said Pooser.

For sure, these are still modest size problems (number of orbitals) but perhaps not for long. Recently HPCwire talked with Pooser about many aspects of the evolving quantum computing landscape – benchmarks; qubit technologies including emerging optical approaches; new cryo-electronic control chips; how the boisterous national quantum community is managing itself; and more. Presented here is small portion of that conversation.

HPCwire: It’s been about a year since we last talked. Let’s start with what’s happening broadly.

Raphael Pooser: There are now these national quantum science centers that formed. This was the culmination of the National Quantum Initiative, which people have been talking about nonstop for the past couple years and came to fruition in 2020. There are five of these things and Oak Ridge got one of them. This whole thing, under the National Quantum Initiative, is a really interesting program because it’s trying to leverage quantum information across more than just quantum computing, but across a huge spectrum of science. For example, there’s a center dedicated to networking, and the transduction of quantum information. There’s another center dedicated to different types of quantum computers trying to build links between [them]. They’ve got superconducting and ion traps. And there’s a few more.

Our center (The Quantum Science Center) is dedicated to almost anything you can think of – except for networking – to do with quantum science.[i] So we’re building quantum computers. We’re building quantum simulators. We’re building quantum sensors. One part of our center is dedicated to trying to detect beyond the Standard Model physics with quantum sensors. So you have that in the same program as trying to build topological quantum computers.

So it’s a huge effort. When you have $25 million a year, you can’t really throw all $25 million behind one problem. I’d say our flagship problem is building topological quantum computers, which does dovetail very nicely with my research, which is why I’m involved in that center. So what I do [now] kind of exists in the context of all of what Oak Ridge is doing in National Quantum Initiative.

HPCwire: How does the benchmarking you’ve been involved with fit in?

Pooser: In my case, all the benchmarking stuff that we’re doing is targeted towards the eventual technologies that will come out of these centers; it’s like a direct handoff in terms of being able to run these [benchmarks] on next generation devices. That’s kind of exciting. But at the same time we’ve got ongoing projects with DOE already, these testbed programs. There was the Quantum Testbed Pathfinder and the Quantum Testbeds for Science[ii], right?  Putting the National Quantum Initiative to the side for a second and talking about these ongoing testbed projects, there are updates there. There are two DOE-run run quantum computer testbeds (Sandia and Berkeley) that got stood up within the last year. These machines came online in 2020 but the reality is it was a very gradual process – they didn’t just flip a switch.

Now, our main goal, probably one of our biggest goals in the Pathfinder program, is to benchmark and understand the performance of these DOE testbeds. It’s of course important for us to understand all quantum computers out there, but the DOE-run testbeds represents a unique opportunity because they can give you close-to-the-metal access that you couldn’t necessarily get elsewhere. That’s not to detract from other quantum computing platforms out there, which have frankly been the source of pretty much all of our papers.

But there is a unique opportunity with these testbeds because we can do actual physics experiments on them to ask, ‘what would happen if I plug this thing in a different way?’ all the way down to, ‘what would happen if I completely changed this waveform?’ where there’s some tolerance for, I don’t want to say breaking, we don’t want to break anything, but there’s some tolerance to push things to their limits, to see where computational capabilities break down. As my counterpart Robin Blume-Kohout at Sandia would say, one of the best ways to find out what a computer is capable of is to break it and then figure out why is it broken. These testbeds can really serve that purpose, which is a kind of counterintuitive way of thinking about it.

HPCwire: Which qubit technologies are you working with?

Pooser: There are two well-understood technologies at this this point [and] they’re the dominating technologies. The one at Sandia is an ion trap device. The other one (at Berkeley) is a superconducting device, actually plural devices. They’ve got an interesting device; it’s a dilution fridge that can house eight different quantum processing chips. Each one, I think, can have up to about eight qubits. So they could, in one dilution fridge, have 64 qubits at any time, but with different physical chips. The idea is they want to test different architectures. They want to have a guy like me say, ‘Here’s a program that we can run on all eight of your different chips and let’s see what each one does, and which one does it best.’ This is where benchmarking and development of the architecture come together in co-design.

Now, the Sandia device is a single chip, but it’s based on sort of the industry-leading ion technology, which is Ytterbium ions. It’s currently a room temperature device, which means that they have a surface of electrodes, and they have a bunch of ions hovering above it, and it is operated at room temperature. They can cool that with a cryogenic system and I think that’s where they’re heading now, and doing that makes everything a little bit less noisy.

HPCwire: What about the optical technology approaches? Are you working with any of them?

Pooser: Yes, we have access to the Xanadu device in Canada. They just came out with an architecture paper where they described running continuous variable quantum computing on their system. This is a system based on the amplitude and phase of the optical fields, and it treats those as the quantum variables. So normally, as with ions, it’s the electron energy level, or it could be the nuclear energy level based on the spin, that gives you the zero or one. On the superconducting qubit, they’ve got a resonant cavity with a microwave photon in or out.

The optical ones are a bit weirder in that they don’t have zeros or ones; they have sort of a continuum of numbers [based on] quantum variables, amplitude and phase. That’s actually one of my domains of expertise. I don’t know if we’ve talked about this before. But I did my Ph.D. in continuous variable quantum information. So we’ve got this access to the Xanadu device, which is quite an interesting device. We also have, as part of the Quantum Science Center [at Oak Ridge] just started an effort to build a continuous variable quantum computer for the purposes of running quantum field theory simulations. Some folks have discovered that the continuous variable platform may be more amenable to certain types of encodings.

The qubit platforms are naturally very amenable to fermionic encoding. So electrons, which means chemistry, work well there. It turns out that bosonic encodings work very well on the continuous variable platform and there are lots of field theory problems that one can explore. We’re pretty excited about it. We haven’t started and are right now just getting the funding. I’ve been working on that field for years, building quantum sensors. So the same technology that lets you build a quantum computer will also let you build a quantum sensor. My main goal in doing all that was to show that quantum sensors and the continuous variable quantum information used to make them could be used to develop the techniques for quantum computing. It’s been a slow process. We have not done a whole lot with continuous variables here aside from theory and aside from quantum sensing.

HPCwire: The optical material is fascinating but less well known. Have you looked PsiQuantum’s technology and is it like Xanadu’s?

Pooser: They’re similar, but different. PsiQuantum is using – instead of continuous variable amplitude and phase – they are using qubits. So you can also cause optical fields to represent a two-level system, and it can become a qubit. There’s many ways you can do that such as polarization or some other variable to discretize the continuous variable if you want. But they are similar in the sense that they rely on similar technologies; that is they have entangling gates which join optical fields together and cause them to be entangled with each other. They create these large resources of entanglement that they then operate on with other quantum gates to generate the results. But that’s about where the similarities end. They’re drastically different devices.

In fact, the approach that PsiQuantum is taking is really interesting. They’re shooting for massive scalability from the start. And Xanadu is taking more of the NISQ approach, which is what most of the companies have taken to date. PsiQuantum, on the contrary, [has said] our only goal is to reach fault tolerance. So that’s the way they’re doing things and there’s nothing wrong with that either. It’s also a good approach.

HPCwire: Right. They are focused on leveraging existing tier 1 fab technologies to have finely defined feature size.

Pooser: Yes. I was really elated, honestly, to hear about it. We had a meeting with them recently, in which they went over their current roadmap, where they are on that roadmap in terms of the technology, and it was pretty amazing the advancements they had made in the fabrication. It was one of these things where years ago you sort of knew what could be done, based on papers that you’d see out there and demonstrations. It was sort of one of those things was like, okay, that’ll work but it hasn’t been tried. Now they’re actually doing it. It’s really hard to tell what the end result of their attempts at scalability will be, but what they showed us was pretty impressive.

HPCwire: Let’s circle back to your benchmarking efforts.

Pooser: So I sent you a paper that we’ve just published (Benchmarking Quantum Chemistry Computations with Variational, Imaginary Time Evolution, and Krylov Space Solver Algorithms, Advanced Quantum Technologies). One of the things that we’ve been doing for quite a while has been trying to test how well current day quantum computers can handle simple applications, and sort of try to use that as a predictor for how well they will handle very complex applications that are based on similar concepts in the future. So we’ve been taking every new algorithm that comes out and just running it and benchmarking the current devices with it.

First, I’ll tell you about these new algorithms, [loosely] quantum imaginary time evolution. They’re all quantum simulation algorithms and they try to simulate some kind of quantum system and understand how it might evolve with time. Even the Variational Quantum Eigensolver, which is not a time dependent simulation algorithm, is very interested in understanding what the end game is in a chemical interaction, the ground state. They’re all interested in the evolution of quantum systems. So, benchmarking them seems like a good idea. Various journals have asked us to contribute to their body of knowledge along those lines. We’ll frequently get questions from chemistry journals asking us if we can extend the benchmark that we’ve come up with in the past to different devices.

The upshot is from what we’ve seen that the quantum computers have been advancing at quite a rapid pace. So two or three years ago, we were seeing that we could work really hard to get interesting results on quantum chemistry out of the quantum computers of the day. And the concept of chemical accuracy, which is sort of the gold standard, was in a nutshell very hard to attain on the hardware if you didn’t have an in-house device that you’d built yourself. Fast forward to today, we just got through running this benchmark on the latest quantum computers of the era from IBM, and we have some unpublished results from other devices, and these systems’ performance has grown by leaps and bounds. It’s gone from being very hard to achieve chemical accuracy on those same problems three years ago to becoming routine now.

So the pace, in my view, is very promising. It means that if this metric, chemical accuracy, if [achieving] it becomes routine for simple problems that we’re working on now, that means now we can think about increasing the complexity of the problems, increasing it up to the point where the accuracy drops again. At that point we’ll know the [new limit for simulating these] more complex chemical systems and other quantum systems.

HPCwire: How does what you’re able to do now compare with what you can do on big classical systems?

Pooser: We’re still a long way away from there. If we’re using chemistry, we need to be able to do on the order of 100 orbitals to calculate ground states of molecules of interest that are hard for classical computers to calculate. Then, of course, there are many other advances on classical computers that quantum has to catch up with, for example, protein folding. People throw machine learning at molecular dynamics problems instead of actually doing molecular dynamic simulations. One idea is let’s throw the theory out the window and try to learn it and it does work.

Quantum computers are far away from being able to do anything like that. One of the reasons is that, while they’re getting very good at quantum simulation, there’s still a fair bit behind on the machine learning. Classical computers currently excel far beyond what quantum computers could do on machine learning. I think one of the reasons for that is because there’s been a lot of focus on quantum simulation in the NISQ era and people have really targeted at making machines perform well on that task. The other reason is that some types of quantum machine learning are just harder. While some types of classical machine learning are just easier for classical computers today.

HPCwire: Where do you get the chips and systems you work with? Do you build them or use those from companies that are currently in the commercial sector?

Pooser: It’s both. For most of the stuff that we’ve done over the past four years in terms of developing algorithms for quantum computers, that has been targeted at industry developed chips. So IBM, IonQ, Honeywell, all of them are commercial entities now. Sometimes it’s blurry. If you look at IonQ, they started out as a lab at the University of Maryland. Our interaction with them here at Oak Ridge started out that way. We were running simulations on their device at the university, but it was through the company IonQ. They had their headquarters actually at the University of Maryland campus, [and] they sort of naturally continued on to a fully-fledged independent economic entity.

HPCwire: Looking at the applications, have there been differences with regard to the underlying qubit technologies that you might comment on?

Pooser: Yes. The different qubit technologies are suited to slightly different tasks, even though it’s the NISQ era, right now. A lot of folks will say that Variational Quantum Eigensolver, for example, runs better on a superconducting device because you can measure quickly. You end up being dominated by the measurement time and, of course, you need to do the same circuit measure many times and that penalizes the ion trap devices a little bit because they have a slower measurement right.

The flip side is that ion trap devices have been demonstrating very good quality quantum machine learning applications. One of the reasons for that is because the interconnectivity of the qubits is higher than on the superconducting technology. If you had a fault tolerant device, it wouldn’t matter because with a fault tolerant device you could engineer any connectivity between qubits, logical qubits, you’d want on a superconducting device. But we don’t have that yet. To make complex neural networks requires physical connectivity which is limited on the superconducting devices. On ion traps there have been demonstrations of interesting quantum machine learning applications coming out of a company called Zapata – at first NASA was publishing this and then the folks who were publishing those benchmarks moved to this company, Zapata in Toronto.

If you look side by side at those applications that they’re running, you’ll see that you get a bit better resource usage on ion trap devices, but it’s not a big enough gulf in any of those spaces to say that it’s a big deal right now. For example, we can do VQE on ion traps, no problem, it’s just a little slow. And we’ve done plenty of machine learning applications on superconductors.

HPCwire: Some of the recent work on cryo-control circuits is interesting. Intel comes to mind. Have you looked at that area?

Pooser: We have. Google, I know, is also focusing on this. I’m pretty sure IBM must be, although we don’t talk to them a lot about that level of hardware, at least work from where I sit – we talk to them a lot about the stack, and about how the stack interfaces with the hardware at the quantum control level. But the answer is yes, we are actually working on this too. So back to National Quantum Initiative, we realized that both superconductors and ion traps, and actually even optical devices can benefit from on-chip, quantum control electronics. In the case of superconducting and ion traps, they should be cryogenic. Luckily, in the optical case, the on-chip controls don’t necessarily have to be cryogenic, although you get benefits if they are. We’re developing cryogenic control electronics for ion trap devices.

We’ve got an in-house device here, that’s complementary to those devices in the sense that it’s an analog quantum simulator. So it’s not doing the same thing as other quantum computers that are out there but it benefits from the similar technologies that you just talked about.

HPCwire: Similar to the D-Wave approach?

Pooser: That’s right. It’s not exactly adiabatic, but we are setting up a chain of analog spins in trying to understand phase transitions and model materials using dynamics [by] observing the dynamics of this chain of ions and understanding how can they stand-in for the Hamiltonians of materials. In partnership with Fermilab, we have an effort to develop the control electronics for that device in an integrated fashion at the cryogenic level. I think it’s complementary to the efforts going on in the superconducting space.

To go further. The other way that we’re trying to use and benefit from this is with IBM which is developing and had already released mid-circuit measurement capability. Sometime in the future will come the ability to use the results of that mid-circuit measurement to actually address the subsequent circuits. In other words, you do a measurement and you say, now what should I do? That’s a key. We aren’t able to do that in quantum computing yet. That’s a key milestone, right? It’s like quantum if-else. We don’t have a quantum if-then yet. Once we have that, that’s the first step towards implementing quantum error correction.

There are a many other things you can do with that, too. You can simplify many algorithms and make them shorter depth. One of the things that IBM realized – and we’ve talked with them about this and we’re actually working on the stack to support this – is that there are a lot of operations you’re going to want to do in those on-chip electronics. Once you have this measurement and feed forward capability, you don’t want to have to propagate all the way back up to the classical computer because that’s going to have a lot of latency. Meanwhile, your qubits are waiting. And they’re like, well, I sure would like to sit down, which means decohere here. So they’re standing around and they’re looking for a seat. If you can get back to them ASAP, you can do more in the amount of time that they’re willing to stand up. If we start implementing quantum kernels — at Oak Ridge we call them quantum kernels, I’m not sure what IBM calls them – but the idea is that if we can implement quantum programming kernels all the way down at the quantum control hardware, and make some decisions down there before you ever have to do anything with the classical computer, you win.

HPCwire: Thanks for your time, Raphael.

Notes

[i]DOE description of ORNL Quantum Science Center: “QSC is dedicated to overcoming key roadblocks in quantum state resilience, controllability, and ultimately scalability of quantum technologies. This goal will be achieved through integration of the discovery, design, and demonstration of revolutionary topological quantum materials, algorithms, and sensors, catalyzing development of disruptive technologies. In addition to the scientific goals, integral to the activities of the QSC are development of the next generation of QIS workforce by creating a rich environment for professional development and close coordination with industry to transition new QIS applications to the private sector.”

[ii] Broadly, DOE’s Quantum Testbed project is a multi-institution effort involving national labs and academia whose mission has two prongs: one – the Quantum Testbed Pathfinder – is intended to assess quantum computing technologies and deliver tools and benchmarks; and the second – the Quantum Testbeds for Science – is intended to provide quantum computing resources to the research community to foster understanding of how to best use quantum computing to advance science.