Integrating Quantum Computers and Classical Supercomputers with Martin Schultz

The New Quantum Era, a podcast by Sebastian Hassinger.

And Kevin Rowney

Welcome back to the podcast. Sebastian here with you for another solo episode. Today, I'm coming to you from the IEEE International Conference on Quantum Computing and Engineering at the Palais des Congres in Montreal, Quebec, Canada. This is a excellent conference, put on by IEEE Quantum. IEEE, of course, being Institute of Electrical and Electronics Engineers, which is a professional association, formed in the sixties, I think, that covers a bunch of different areas within the field of computer science, computer architectures, system architectures, electronics engineering, that sort of thing.

So the Munich Quantum Valley is a is a large, effort across multiple institutes, multiple universities. So it includes Fraunhofer Max Planck, TU Munich, where where I'm located, the, Ludwig Maximilians University, the ELLEN University, DLR, and then also the Bavarian Academy of Sciences, which includes the the compute center as well in Munich. And so this is a large, large effort that was born out of, the Bavarian high-tech agenda that's pushed by the state. And they kind of wanted to establish the quantum science in Bavaria with a very large grant that was then supplemented with other grants, also from the federal level, from the EU level. So it's a lot of stuff that that came together, and it started from the physics side, like most quantum projects started, with the idea of having multi modalities being developed in in Munich or the Munich area, and pushed it up.

the beginning.

And what came clear very, very quickly that we need more

Right. Than

that. And so we started building a software stack, connecting it to, to to a web portal that can we can access things that we know for most quantum systems. But then very early on, we already thought, well, this is the web portal alone won't help us that much. We need to do this in a way that we can, support a wider workload, and quantum will be always very specific. I mean, we had a conference.

Need

to connect it into high performance computing, HPC. Right. We heard this morning also in another keynote a lot of connection also to AI. Right. And then how can we tie these three things together and make them really work together?

Right.

So quantum Quantum and HPC integration was high on our agenda from the very beginning on. We started with activities there very early on with MOF sessions at various conferences. We have a workshop now that's in its 4th instance on bringing people together. Right. Also, that started, of course, very slow because it's not it wasn't people's forefront in the beginning Right. But it's coming more and more. Yeah. And we're really seeing these communities grow together, which is kind of exciting.

That's very exciting. And do you think in part that perspective came from I mean, LRC is a supercomputing resource that's, very heavily used by the academic community, as well as in collaboration with Industrial R and D, I think as well. Right? There's sort of industrial collaboration projects that go on it?

So it has. So it has a very, very strong and strict public mission at the moment. So the projects have to be public and basic research. Industry can we're working with industry in collaborative projects. It has to be always an academic partner with it. It has to be a funded project from the outside. But then we have some industry collaborations Right. Of as well. That's actually one of the challenges we're looking at right now. How can we open this up to have more industry involved?

know? No. I mean, it's what I find fascinating is, you know, you have sort of you're thinking over the horizon of, getting these QPUs to the state that they can actually do useful work. And you're thinking about what does that solution what does that end to end solution look like? And, you know, I think you're not alone in that.

Right. And so we we we actually managed to do a first shot between the HPC system and, the the quantum so one of the quantum systems in direct, in in direct connection there. And so we have been getting to this with, kinda, we we have this concept as our quantum team calls this as a data series of hero runs. Mhmm. So we have, like, these goals, these goal posts that they're trying to reach and and and really really shooting for, and they manage to do this connection between the HPC system and the quantum system in a very additional thing.

Yeah.

But, yes, we've been trying to do this for, we've been working towards this goal for for a while, and to be honest, also, the the emphasis has shifted. Right? In the very beginning, it was, how can you take a system and put it in a data center? Right? How do you even hook it up? Right. Or how do you What kind of environment do you need? How do you deal with the false floors which we have in centers, which is terrible for quantum systems because there's Lots of vibration. Leverage.

Exactly. The opposite of what, what a classical system is about.

Exactly. And so also the, and also then even if you have it on a floor, on a on a solid floor, how do you put in together then the the right water? They need slightly different water water chemistry, different water temperatures. How how they get the right power together, how they hook the the the networking together. We we focus at the beginning very much on the on the hardware side just to get this thing going, and then we had some duct tape or software in between them just to make it work.

That's such a valuable set of experiences, and really, there's no substitute for putting one of those devices together. They're they're kind of a magical combination of science and technology right now.

Right. Exactly. It's really tough. And so we learned a lot in in this part of the project, and by now, the more newer systems we, quote unquote, buy. The hardware, of course, is not that easy. It's not like Yeah. You buy a rack, plug it in, and then run it. Yeah. But there's still there's more like of of a of a consumer seller relationship. Right.

Is that from an existing vendor, or is that a

Yes. There's an there's a system from from AQT, Alpine Quantum Computing, which are very close partners with the Right. The new quantum valley. I mean, they're just hour and a half away. So this is kind of the nice thing in the sense.

And are these systems sort of you call them r and d systems. Are are you is there sort of, whatever access into the the system itself? So, you know, some vendors, their system might be treated as a black box and you sort of you can send your circuits and get your results, but you don't really know what's going on inside. Is this more of an open system to sort of mirror the open software, development efforts that you're doing as well?

So that's a that's a little bit longer story. So it's a yes or no. But like it always is. Right? Yeah.

I see.

I see. And that is kind of a a trade off for now because that still leaves a lot of work to do on the vendor side, and then we have our stack on top of that.

Right.

And the idea now is to work with the vendors to slowly migrate these heavy vendors to act into something very thin and low level and push this over to the other side.

Interesting.

And that stuff is working. A lot of these contracts that we have, when we buy these systems, have in there that this transition has to happen. And, you know, the vendors we're working with are very open for this. We have regular conversations on this. They're they're interested in having an interface into the compute center that is standardized.

I always say, use your PhDs to work on the cubits. That's the hard part, not the classical software stack. We have people that do that. Those are solve problems.

Well, it's also the hard part. Right? But but in fact, for for a different group of people, you're absolutely right.

This is a different expertise there. Businesses writing code.

I won't comment on that.

Yeah. So are there existing frameworks or code bases that you've been leveraging? I mean, I know, you know, Kik, Cubic, Kibo, those are all sort of, they have control platform, open source control platforms, but also Kibo's got a whole set of open source libraries. There's, you know, there's there's open source projects from, the Unitary Fund, for example. So there is sort of a it's a small but, growing community of existing open source technologies.

Bits and pieces. So we we're starting to bring things in where it makes sense. The the base infrastructure that we have, we're actually leveraging open projects from the HPC community or from the classic compute community.

Slurm?

We we we do leverage Slurm. That's a separate story we can come back to. We're thinking whether Slurm is sufficient for us or whether we need some more or not.

I just like to say the word Slurm. It's a funny funny word. I don't know.

Well, I don't know if you like this. The other ones we're looking in is a is a skater called Flux Mhmm. Which is out of, out of the DOE Labs. Okay. Livermore. Yeah. So Flux or Slurm? We'll see. Yeah. But, yeah, we we need to do more more more scheduling approach. It kind of allows us to do this integrated scheduling, but that's an open question of where where we wanna go. And there, we're watching very closely other software projects on the pure HPC side

Right.

That need similar techniques, similar software components. And we we don't wanna do anything separate. We wanna kind of fit into the ecosystem. Yeah. Yeah. Yeah. So there's a tight relationship there to other software projects as well

to make this work. I mean, things like, scheduling, those tend to be more batch or, you know, less performant, dependent, sort of like it's queuing up and using a resource when it becomes available. You also mentioned AI, and also, obviously, the the whole talk of, of state preparation or error mitigation or error correction, all of those need, I think, really, really high performance, and tight integration between classical and and quantum. So you don't incur overheads that eliminate whatever kind of advantage you might get from the quantum algorithm. What what sort of, integration strategies are you do you have in mind for those sort of tighter I mean, more demanding kind of use cases?

Yeah. So there's there's gonna be different time levels and and and time loops, basically. As you said, there's just different granularities that that we need to have. And so from the the very coarse grain, we're gonna start with the job scheduling systems where you said you want a quantum system. You wanna have access to that, use prediction mechanisms that allow us to predict when resources are free and how can we overlap Right.

Right.

And the question is how much we can, how tight are these loops. Whereas in some technologies like superconducting, we may not be able to do this purely in software. You have to push things in in hardware because the loops are so tight. Yeah.

Yeah. Yeah.

Other technologies like ion trap, they're a little bit slower and have other qualities instead of that, and then we may be able to bring it into software.

Right.

And our goal is to develop the software stack in a way that it actually can run on the HPC system itself, and connect the HPC system directly to the control.

Like making RPC calls or something too.

Perhaps even lower weight than that to really directly drive this from there. And then you can actually run a feedback loop directly into your HPC system, where you have the resources. We also can use GPU acceleration or FPGA acceleration on on the node itself. Interesting. And for signals like Iron Trap, I think this will allow us to bring us a lot of stuff in.

And neutral atoms as well.

Neutral atoms as well. Exactly.

The lower clock speeds. That makes a lot of sense. But you're right. Something like superconducting, you know, you're probably looking at something like, you know, quantum machines DGX or whatever, where there's a GPU that's actually part of the control system that can do things like error correction at a speed that makes sense for for that.

But even then, you may need some other intermediate loop, kind of, for some kind of conditional executions, and things like that that you want between the the device selection and this low level part. So we have, potentially, another level of of scheduling in between here. So this hierarchical scheduling that really informs from the all over from the grip from the coarse grain to the fine grain and has this this common overview, I think, will be really a key technology. And that some of our our teams are working on that right now.

That's really great. How how large is the effort? How many people are are actually writing writing code with you?

So this is, again, a a very difficult question to fully ascertain. The whole quantum valley is about, 3, 350 people working on it, but that includes all the physics parts Yeah. As well. The software in total, I would say, is about 30, 35 LTEs.

Significant Yeah. Open source effort.

It's and there's a large part in the, in the compute center at LRC. There's a special team, the, the quantum computing department, which has these people and which work directly on the machines and which are also a little bit more geared towards the software development. And then we're working with the university side, we're a little bit more the research y side is. So we're finding this avenue of how to take research, harden it, and then and then put it put it into the stack.

And is there any thinking about, like, integration of cloud based systems, for example, for, you know I mean, in classical HPC, there's an idea of bursting out to the cloud if you need additional capacity temporarily. Is there sort of a similar kind of thinking around maybe a specialized QPU for some aspect of the of a of a workload in the future, obviously, when these things can do productive work? Is there kind of that that thinking as well, that that, remote, or cloud based resource?

That is possible. So the FOBSTA stack itself as well, it'll be flexible as modular as you can plug in different components. And at the end, there's a series of back ends. Mhmm. And the it is also not only intended for us in Munich.

That's great. Yeah. I'm just imagining because, you know, I think that we're going to see probably more diversity in hardware efforts in the future rather than less because we're still in this very exploratory stage. So the idea of being able to stand up an experimental qubit or experimental system in a university somewhere and still be able to use it in that integrated kind of of environment is, very powerful, I think.

So we're actually using this also inside the Munich Quantum Valley, because we said we're developing also the the the the Valley is developing also these these computers. So the Max Planck Institute Right. For quantum opsing is doing neutral atoms. And then there's the Walter Meissner Institute, who's doing superconducting qubits and also experimental qubits there as well. Yeah.

Exactly. Yeah.

And those will never be at the compute center because they need it in their labs. Exactly. And so we have already in the stack now to dose 1 connection. We we demonstrated this from the HPC system going to, a small demonstrate at Walter Meissner as well Excellent. Across campus that works. Excellent. Don't get all the benefits of the tight loop and integration, but it works in therefore responsible.

About the efficiency. Right? If you can if you can, not have to replicate all the work you've done to do those integrations and just tie into it. It doesn't matter necessarily that you don't get the performance as long as you're being able to reuse those tools and those frameworks. It's just shortening the time to actually get it into the hands of the people you want to experiment with it.

So we're thinking about it's doing relatively soon. We have a first repository open. To be honest, this is not one of the most exciting repositories yet, but it was a starting point.

Yeah.

But we have the idea is to open this up in the in the coming months, I would say.

That's exciting.

Particularly focusing first on some of the interfaces that we need. We have an interface definition that works for the vendors, how to hook in. We're working currently with selected vendors to get first feedback, and this is one of the first things we're gonna open up to get broader feedback. Yeah. Perhaps even kind of a sort of a standardization going around it or homogenization, let's say, this way.

That's what I wanted to ask you is is how do you balance I mean, you need standardization in order to make these types of frameworks work, frankly. Right? I mean, you have to agree on standards, in order to be able to make systems talk to one another. But we're also in this very, early stage where there seems to be far more unknowns than knowns at this stage in quantum computing. So is there is there something in your, architectural design that's sort of baking in that uncertainty?

So we're trying to do that. I mean, is this prediction of the future? Why do we don't know? But a lot of us were involved in large software efforts and standard softwares in the past, where we had we've seen how this evolved over time. So, personally, I'm involved in the message passing interface standard, the MPI standard, and I've been leading the the MPI forum for for a while now.

That's great.

But then also, to be honest, right now, we are not yet afraid of breaking things as long as we don't break it too much.

Yeah. Yeah.

So there's still the we won't release a 1.0 version very much on purpose. It will be a 0.1 or 0.2. Yeah. And by the time you get to a 10, that's where we're gonna start freezing and have a more formal change process in place. But right now, I think it's more of an experimental thing. We're getting feedback from everybody, and we're baking this in stone.

That's smart. That's smart because, I mean, I think, you know, there's potentially more diversity coming down the road in terms of the the execution models and maybe in terms of the intermediate representations. And and certainly, everyone has sort of a general assumption that we're going to be climbing up the abstraction stack. So it may end up not, you know, passing around, gates and circuits, but something higher level abstractions that are are actually how you describe your workloads. Right?

I think. And also for the intermediate representation, we're trying to work here also very much with the community in looking into into, standards that exist that have started coming up. So LSE joined, the QIR consortium Yeah. To kind of see there, well, how can we extend this? How can we make this use?

That's another topic, compiling and transpiling. I mean, there's so many differences between it feels like to get optimum performance from any particular piece of hardware, you have to be very, very, conversant in the exact idiosyncrasies of that hardware.

Right.

Is your compiler sort of also architected to to, to be extensible so you can have extensions for their their hardware specific? Yeah. Okay.

Yeah. Absolutely. It has to be this way.

Yeah.

So what we are trying to do, and what I think we have, achieved is a core infrastructure

Mhmm.

That is completely agnostic of the particular technology in the back, which allows you to basically take a circuit, trans get a circuit through a pipeline, change things on the fly, and at the end, have this interface to the back end part. To be honest, it's not even particularly not even quantum specific. Right? This is a disaggregated accelerator

Mhmm.

With just in time compilation Mhmm. Mhmm. With something on the computer science side, we have had Right. Other examples for them trying to retrofit on the data.

Ways in which we can show the physics community that we we're not, you know, we may not get some of the stuff you're talking about. We're pretty smart too.

It has to be a give and take. Right? I mean, we have to have to work together in this one. Exactly. And this is perhaps even the been the most fascinating things over the last 2 years. If you can come back to the the software in a segment, I just wanna put this in here. In the beginning, when the Quantum Valley started, it was where the the physics community and kind of us on the computer science community trying to talk to each other. Mhmm. Mhmm. There was totally different languages.

That's the part that absolutely keeps me completely fixated on this this, this industry, because it's exactly that. I mean, it's it's, you know, not only have we rewound the clock to, you know, hand setting registers, like, we're we're right down at the metal. So we get to repeat the entire evolution

Yeah.

Of classical computing, which is fascinating in and of itself. It's like, what if we did classical computing over again, but we had, the cloud and open source software and higher level languages already. Right?

Exactly.

And then added to that, we also get to have this dialogue with the science community, the physical sciences, in and, you know, opening their eyes to what we can do to help them in along this journey. It's really it's a fantastic set of of challenges and and, you know, sort of intellectual, puzzles to solve.

Absolutely. It's also why I like this this conference where we are right now, the the Quantum Week, which has exactly this is not a physics conference. It's not a computer science conference, but it really brings people together. And I think, also, that helps establish this common language and the cooperation they need.

Exactly. You know, and I think the IEEE is very smart. I've seen this conference evolving over the last, whatever it is, 4 years now. It is exactly that. It's carving out that middle space and now, you know, birds of a feather, sessions of the type you described with, you know, Oak Ridge or, other supercomputing centers, you know, all collaborating around these discussions are now leading to, as you said, potentially early standardization discussions, which will just take more of the friction out and more, you know, enable more, innovation on top of the work that's going on.

Exactly. Yeah. Yeah.

That's fantastic. So, back to the birds of a feather thing. Do you, do you have sort of a first, international collaborator or partners that, you know, like an Oak Ridge or another center out there that that is, sort of waiting for the release of your software to try to join in to the collaboration?

So we definitely have interest from partners to to look into this, or most particularly to work on the interfaces with us. Right. So we've been talking to to a lot of partners, including in the DOE, but also in the German ecosystem. Also, the European ecosystem, there's a lot of effort going on

here as well. PC. Yeah. Yeah.

On how to homogenize this. Particularly in EuroHPC, there's this and as you know, this program of these first 6, Swanon Computers being seeded by EuroHPC, one will be in Munich in the Leipmann supercomputing center. The tender is still out, so we don't know yet what it's gonna be, but it will be, what we know as superconducting system as well. But there's 5 other centers, and the European team made it very clear we want one software stack, not 6. Yeah.

We gotta get, Iced involved from Japan as well. They're building out their g quads.

Yeah. So we have been talking to the Japanese community as well. They have been a part of our work. We need to make this international. Right?

Absolutely.

And for this working group, the 5 of us are really just the starting point. Yeah. And we wanna broaden this out, and so that's why the BoF session on Thursday here, hopefully, will help to bring in more people. Absolutely. And the idea is, at some point, also start your own own workshop series on this to kind of bring more and more people in. Yeah.

I'll be there. Excellent. Wonderful. Well, thank you very much, Frank. I think this is a excellent effort. And I'm really impressed, as I said, when I was at the tutorial last year. And the ideas were great, but I had no idea you're gonna make this much progress in 1 year. So very impressive.

It was a very busy year, but it was a lot of fun.

Yes. That's great. Thank you very much.

Okay. That's it for this episode of The New Quantum Era, a podcast by Sebastian Hassinger and Kevin Rowney. Our cool theme music was composed and played by Omar Costa Hamido. Production work is done by our wonderful team over at Podfi. If you are at all like us and enjoy this rich, deep, and interesting topic, please subscribe to our podcast on whichever platform you may stream from.