Sam Stanwyck, Group Product Manager, Quantum Computing at NVIDIA is interviewed by Yuval Boger. Sam and Yuval talk about NVIDIA’s QODA library, classical/quantum integration at HPC centers, and much more


Yuval Boger: Hello, Sam. Thank you for joining me today.

Sam Stanwyck: Hi, Yuval. Great to be here.

Yuval: Great to have you. So who are you and what do you do?

Sam: My name is Sam Stanwyck, and I lead the Quantum Computing product team at NVIDIA. So I’ve been doing that for about a year and a half now. My background is in experimental condensed matter physics. That was my grad school work. And then, I worked for a quantum computing startup called Rigetti Computing for about three and a half years, and then worked on the quantum computing team at Keysight Technologies, working on quantum control systems for about a year, and now at NVIDIA.

Yuval: Excellent. We spoke last year, and at that time NVIDIA was just announcing some of their quantum products. What’s new with NVIDIA and quantum since we last spoke?

Sam: Yeah, a lot’s new. So I think last year we talked about cuQuantum, which is the set of libraries that NVIDIA released for accelerating quantum circuit simulation on GPUs. And the goal there was to really enable algorithms research. So probably, as most people know, today’s quantum processors are pretty limited in scale and especially in error rate. And so if you want to research and develop new quantum algorithms, test out their sensitivity to noise, try to figure out for a given application when quantum computing might be useful, simulation is a really important tool. And the GPUs that we make are very, very good for simulating quantum circuits. And so cuQuantum was released a little over a year ago now, and our latest effort is kind of broadening the scope from that into thinking about integrated quantum and classical computing. So I think most listeners of this podcast will also know that as good as quantum computers get, there’ll be plenty of things that they’ll never do well just because of the kind of limits inherent to using quantum physics to do computing.

And so what accelerated systems based on quantum computing are going to look like in the future is quantum computers tightly coupled to classical computers, so CPUs and GPUs and QPUs all working together and in close communication and each one doing what it does best. And for that to work well and for a programmer or a researcher to be able to really take advantage of that, we need software that enables using all of these together in a flexible way and in a performant way. And so last summer, we announced software called QODA. QODA stands for quantum optimized device architecture with the goal of enabling easy and high performance co-programming of quantum computers together with classical computers, CPUs and GPUs.

Yuval: Does that focus primarily on orchestration, making sure that the classical part finishes and then send off the data to someplace and then get it back from quantum and so on? Or does it focus on taking a goal like optimization and then somehow automagically dividing it between quantum and classical resources?

Sam: Good question. So it’s a bit different from and complimentary to tools for orchestration. So QODA is a programming model, and if you’re familiar with CUDA, which is something that NVIDIA has provided for a long time, the analogy is what CUDA is to GPUs, QODA is for QPUs. So you can think of QODA as a superset of either C++ or Python. So you can do everything you can do normally in C++ or Python, including calling CPUs and GPUs and taking advantage of accelerated software to run things on them. Plus, you can add quantum kernels. So quantum kernels are denoted and compiled separately and orchestrated then together with the classical kernels and can be run either on simulated quantum processors or on physical quantum processors. So it’s really a programming model to enable high-level co-programming of quantum computers with classical computers.

Yuval: Do the quantum kernels support only gate-based machines, or are you also looking at annealers or analog machines, other types of quantum machines?

Sam: Yeah, good question. So we think it’s really important, especially at this point, to support different models of quantum computing, not just gate based. There’s such a wealth out there, from more continuous variable type approaches that you might see with a photonic quantum computer to things like analog Hamiltonian simulation that you might want to do with a neutral atom quantum computer all the way to quantum yielding. And one thing that we did at the very beginning with QODA is establish partnerships with leading companies, building different types of quantum hardware.

So that would be superconducting qubits, spin qubits, photonics neutral atoms, and trapped ions. And we intend to keep extending that, but we’re working very closely with those companies to ensure that the model at a high level supports all these different types of processors as first-class citizens. And now you probably will know that compiling circuits for these different types of processors looks very different, and it will look very different in QODA, and for some of them, it will be more of interoperability than the kind of gate model compilation we might do. But at a high level in the model, they’ll all be supported.

Yuval: One of the things that customers sometimes talk about is finding the best quantum computer or the best classical quantum combination for a particular problem that they have. How does your environment help in that? How easy it is to go from one type of machine to another, from simulator to actual hardware, and so on?

Sam: We think that that’s really important work and really important to understand and something that’s really hard to do today. So most quantum computers are programmed in the quantum equivalent of assembly code. So talking about individual gates and individual qubit rotations, and that is needed for the work today, but it’s not so easy to integrate with, for example, a machine learning workflow, say, a generative adversarial network, where you want one side of it to be classical on the other side of it to be quantum, you need to learn a new language and figure out how to join those two, and you’ll likely run into system and compilation level bottlenecks that’ll make it really hard to evaluate if you’re actually getting an advantage.

And so the idea with QODA is that you could really just start from an optimized classical computing workflow and ask the question, is quantum going to be valuable for this? And with QODA, just take part of that workflow or all of that workflow and test porting it to a simulated quantum computer or a real quantum computer or just do a resource estimation. So we expect that it’s going to be a really valuable tool for people looking to answer those types of questions.

Yuval: When you think about enterprise users that use GPUs today, they could use them on a cloud like on AWS, or they sometimes have their own high-performance computing center. How do you see quantum fitting into that? What do you advise your customers? How should they physically get quantum involved in these workflows?

Sam: We expect that long-term valuable quantum computing is going to look like quantum computers and classical computers tightly coupled physically. And so that is something that could be offered by a cloud service provider or could exist as someone’s on-prem system. With that said, I think for the kind of work that’s going on and needs to go on today, it’s perfectly fine to have classical computing resources and quantum computing resources that are not co-located. You’re still able to understand a lot of things and develop new algorithms if you just often hit the communication latency bottleneck that long term is going to prevent you from doing valuable quantum computing. But you’re not able to do that anyway right now, so we think it’s still a useful paradigm for a long time.

Yuval: Could you share a couple of examples of customers using QODA today and what they’re doing with it?

Sam: So QODA is actually not released publicly yet, so look out for that a little bit later this year. But we have announced a few partners. I can’t talk in detail about what they’re doing, but one group is QPU providers. And so what they’re doing is connecting their hardware to QODA and working with us to make sure the model is optimized for their hardware. And so that group right now includes Quantinuum, IQM, Pasqal, Xanadu, Rigetti, and Quantum Brilliance. And again, we intend to expand that. We’ve also announced collaborations with some companies developing quantum software and applications who are looking to build on top of QODA as a platform for that. That group includes Zapata, QCWare, Classiq, and Qubit Pharmaceuticals. And then the last group of partners we’ve announced publicly is a set of supercomputing centers who are either already or thinking about starting to integrate physical quantum computers with classical supercomputers. And that group includes Oak Ridge, Julich, and RICON.

Yuval: If I understand CUDA, the classical part, correctly, then NVIDIA engineers have worked really hard to make sure that I could run an algorithm in a fairly optimized way regardless of what NVIDIA GPU I’m using. Do you see the same happening in quantum, or do you think it’s still very far away and every manufacturer is going to optimize the specific code to their particular hardware and you’ll just be, I think you called it quantum kernels, you’ll just be integrating these quantum kernels together?

Sam: So we think that there’s some opportunity for that now, but when CUDA came out, GPU’s were already kind of a mature hardware technology that was delivering value in production. And we’re not quite there yet with QPUs. So it is a bit early maybe for standardization and a lot of optimization that we would do would be premature. So we do plan to offer some of that, but the main plan for the moment is to open source QODA and work with hardware partners to build in optimizations for their platforms and work with and allow the community to contribute and build in further optimization.

Yuval: Do you see customers looking for essentially shrink-wrapped applications, “just solve my chemistry problem?” Or do you see them looking more for an API that they can do whatever they want classically and quantumly?

Sam: I think more the latter at this point. If there’s a kind of just direct set of steps to use a quantum computer to solve a chemistry problem, at this point, I’m not aware of it. It’s more of the research and experimentation phase. So we tried to build QODA to be very low-level and flexible, right? It’s just anything you can do and C++ or Python you can do in QODA, plus, you can define bits of code that get sent to quantum processors, and that was a big reason why.

Yuval: You mentioned resource estimation. Does QODA also take care of estimating the cost of running an algorithm here or there in some combination of classical and quantum?

Sam: So, resource estimation is on the roadmap for QODA. Of course, it’s not released yet, so it’s not available yet. But yes, in general, we think one of the most important things people need to be doing now and can be doing now to prepare for useful quantum computing is to take the hard computational problems that they care about where classical computing might struggle and do simulation and do resource estimation and understand how many qubits do we expect we need for a quantum advantage? What kind of error rates, and how many clock cycles do we need to run to expect a quantum advantage? And so we’re going to try to enable that in QODA.

Yuval: As we get close to the end of our conversation. I wanted to ask you a hypothetical. So if you could have dinner with any of their quantum greats, dead or alive, who would that be?

Sam: Oh, man. Save the hardest question for last. I guess the first person that comes to mind is Feynman. So in addition to being incredibly brilliant in a lot of different domains, he was, as many people know, very kind of charismatic and interesting to listen to and a great teacher. And he was one of the very early people who talked about what a quantum computer might do as far as simulating nature. So I would think he’d be pretty interested and excited to see how far things have come in the last few decades, and I’d be really interested to get his perspective on it.

Yuval: Perfect. How can people get involved with QODA, learn about it, and perhaps get eventually trained on using the new offering?

Sam: Good question. So we have a website. I may butcher the URL, but I think it’s QODA where you can apply for early access. If that URL is not right, I think just searching for it should bring it up. You can also feel free to reach out to me on LinkedIn. And look out for an announcement of a public beta later this year.

Yuval: Excellent. Thank you, Sam. Thank you so much for joining me today.

Sam: Thanks, Yuval. It was great. Great to be here.

Yuval Boger is an executive working at the intersection of quantum technology and business. Known as the “Superposition Guy” as well as the original “Qubit Guy,” he can be reached on LinkedIn or at this email.

March 20, 2023