Matthijs Rijlaarsdam, co-founder and CEO of QuantWare, joins Yuval Boger to discuss how his company is enabling the scaling of quantum processors through its VIO technology, a 3D chip architecture that addresses key bottlenecks in scaling qubits. Matthijs explains the three major challenges—fan-out, integration of components, and yield—and how VIO solves these to facilitate quantum chips with millions of qubits. He also shares insights into Quantware’s business model, which spans fully packaged chips, foundry services, and packaging services, drawing parallels to the semiconductor industry. They explore the impact of heat, signal routing, and error correction, the strategic importance of marketing and storytelling in the quantum industry, and more.
Transcript
Yuval Boger: Hello Matthijs, and thank you for joining me today.
Matthijs Rijlaarsdam: Thanks so much for having me.
Yuval: So who are you and what do you do?
Matthijs: So my name is Matthijs Rijlaarsdam, I’m co-founder and CEO of QuantWare, and we are a company that makes scaling technologies, a technology called VIO, for quantum processors. So basically we create 3D chip architectures to make quantum processors really large, much larger than you see today.
Yuval: Could I buy a quantum processor from you or do I come to you with a processor and you somehow connect some of them together?
Matthijs: Actually the answer is both. So we make this VIO, which we haven’t been very public about, but let’s say about 70-80% of the people at our company, which is by now a team of about 50 people, growing to 100 next year, they work on VIO. And the way we bring VIO to market is in three product verticals. One is fully packaged chips, so indeed you can go to our website and order a quantum processor. The second is foundry services, so if you’re a customer and you have a great design for a chip, you can come to us and we’ll fabricate it, and because we have VIO, we can fabricate chips much larger than you otherwise would be able to get. And the third option indeed is like you say, it’s basically packaging services, so a customer that is able to make quantum chips comes to us and we package it in VIO and therefore again they can make much larger chips than they otherwise would be able to.
Yuval: How much larger is much larger?
Matthijs: Well, without revealing our non-public roadmap, today we can do devices, like the latest platform we have out is called VIO 176, so that’s on the order of 176 signal lines, so depending on your architecture about 100 to 200 qubits. But what VIO really gives our customers is the fastest path to go to a million qubits in a processor. And so without going into specifics when we think that will happen, that’s a lot sooner than the other roadmaps in the field. And the way we think that is important is that if you cannot make very large quantum processors, you need to resort to networking smaller nodes, smaller processors together, and that will work and networking will be an important part of utility scale systems, but networking also has, you know, transduction has high losses, the network itself has low fidelities and so we think that it will be important to have these nodes as large as possible, and so that is what we focus on.
Yuval: If I think about other superconducting qubit manufacturers, they indeed talk about networking. What is preventing them from building a chip with say 5,000 qubits on a single chip? What do you think is the limiting factor there?
Matthijs: So if you think about scaling superconducting chips, there are three main bottlenecks and you named 5,000, there are some let’s say intermediate solutions that you could probably go to about 5,000. I mean you’ve seen let’s say IBM’s roadmap where they do go to, you know, that number depending on the architecture that they choose. But those three bottlenecks, they are one, fan-out at the qubit chip level. And so what that means is in two-dimensional chips, you need to route the signal to the edges of the chip in order to create a little what’s called a wire bond, so a little connection to the printed circuit board and then to the outside world. And that doesn’t scale for a variety of reasons, I mean just practically the signals traveling on the chip are analog, you want a lot of distance between them if they’re too close together you get crosstalk, I mean that’s one. But also just very fundamentally from a geometric perspective, if you think about the edges of a chip are one-dimensional, the area of a chip is two-dimensional and that means that as you increase the number of qubits you will run out of space at the edges to route your lines. And the problem actually is a little bit worse than that because you’ll get crossovers over these lines etc. etc. So even for our 25-qubit chip Contralto that you can buy on our website, about 80% of the chip space is used to route lines, not for the qubits.
But that’s the first problem and so this means you need to in order to solve this you need to come in from the top. And you can visualize this because the fan-out gets worse than outside of the chip with really large printed circuit boards and so you see these pictures of people holding 200 qubit quantum processors and they really need to put their hands quite far apart. You can imagine going to a million qubits you’ll need printed circuit boards the size of soccer stadiums. It’s not going to work.
The second problem is integration of up-side components. Integrated circuits for classical computers have been around for 60 years, probably even longer but in quantum computing, particularly in superconducting quantum computing, we have the golden chandelier completely full of golden little separate boxes and those boxes contain signal conditioning components like filters and amplifiers and those are separate superconducting chips in separate little metal boxes and they look pretty small but if you do the calculation even at today’s small sizes a couple of grams per amplifier if you go to a million qubits you need to put hundreds of thousands of kilos of equipment in your fridge and so all of that will simply not scale and so you need to integrate that into your QPU and that’s the second thing that VIO does. It’s a fully silicon-based architecture and so we have a lot of real estate there to integrate all of those components because it’s all based on the same materials.
And then the third thing that we that you need to solve is yields and also this is just you know pretty simple like eventually fabrication processes should be stable enough that you can make a million qubits with high probability but that scales exponentially bad so if you have 99% probability that your qubit will work which sounds pretty good but if you that means if you make a chip of 100 qubits the probability that all those qubits will work is zero and so you’re because it’s you know 99% to the power n where n is the number of qubits so the scale is exponentially bad and so the you know one of the ways to break that is to make a large chip out of many smaller chiplets and so you need to solve all three of these problems and there are many intermediate solutions that you know maybe go to a thousand qubits ish and and you see some players on their road maps like putting many chips together in one dimension because then you can kind of get around the fan out or you push it out a little bit further. And you also see players with these blocks of aluminum metal where you put your chips but all of that is either hard to produce or still very bulky or you can’t integrate those components so if you want to go to million qubits this is really something you need to solve.
Yuval: And how does your technology solve these three problems?
Matthijs: Without revealing our secret sauce and since we’re friends and no one’s listening, we will reveal more on this in the coming year or so, but but summarizing a little bit, we create what’s basically a 3D chip architecture. We have a clever way of putting together many silicon chips and because it’s silicon-based we put together many silicon chips and bring in the signal from the top towards the qubit plane so it’s three-dimensional signal routing. We can integrate the components in this 3D chip architecture. Because it’s silicon-based it’s the same type of stuff that the qubit plane is made of and it takes in chiplets. And so the combination of those three things means that it’s a very scalable platform if you will.
Yuval: How about heat? Is that a problem?
Matthijs: So I mean in general the heat is a challenge if you want to scale and that’s because of two types of heat loads in a fridge. One is active heat load so it’s the heat generated by the components themselves. That is fairly minimal so if we do those projections that will not be a challenge to scale to a million qubits. And the second one is passive heat loads that is the heat that is conducted by all the elements on the signal carrying line cable going from the coldest bit of the fridge where the QPU is to room temperature. hat’s as isolated as you can make it but it will conduct some heat and that is actually one of the main problems and so there are a number of clever things we are doing there that alleviate that problem far enough that you can at least get a million qubits in and at that moment again because it’s silicon-based you can integrate things like on-chip control. And also there without going into specifics there are a couple of steps there. One is creating an interface that is much smaller and that’s where also the signal carrying lines simply have a lot smaller footprint and therefore conduct less heat that already pushes it out to very large qubit counts. Then there are a few other clever tricks that we do that make this more of an infrastructure problem rather than a science problem at that moment.
Yuval: When you spoke about fan-out you mentioned the 1D versus 2D problem that the edge is essentially one-dimensional and the surface before your architecture is two-dimensional and let’s even assume you solved that but if every qubit needs two or three control wires then a million qubits still need two or three million control wires how do you solve that problem the number of wires I mean if I look at a Intel CPU it may have thousands of wires or thousands of leads and billions of transistors so how do you get around the million wires problem?
Matthijs: So that’s exactly what our architecture solves. In fact the architecture that we will most likely use what we’re designing around it has more than four million lines. Since we route in three dimensions with all the signal shielding that we do there, that’s exactly the problem that we solve because indeed you need to solve that. And then you have these things like the SFQ that using classical superconducting electronics generates the analog signals from digital signals coming in and then you can multiplex. So in the very long term that also makes the heat load in the fridge less. nd we have a long-standing partnership with SEEQC on innovation like that. There are various other things there so that’s one step further away than where VIO is positioned. What VIO solves is exactly the problem that you sketch getting millions of lines to the qubit plane.
Yuval: What is the status of the technology? You mentioned that one could buy a chip off your website. Who is your customer, who are your customers, or what are the types of customers you’re looking to work with?
Matthijs: So it’s important to in terms of customers to make a distinction of the majority of our customers today they are customers of our fully packaged processors and so those are system integrators, most of them commercial. We currently have customers in about 19 countries so we’re fairly international. In terms of volume we are one of the largest providers of of quantum chips. With VIO we are we’ve just started to roll out to a few selective customers and that’s the VIO 176 platform that I told you about. Because VIO is modular expect that number to go up very rapidly in the coming years. And then we also expect that our emphasis on the three product verticals, fully packaged chips, foundry services, and packaging services, the emphasis right now is on fully packaged chips, we expect that to shift to foundry services and packaging services because the key differentiator of QuantWare is VIO and if you use VIO you’ll be able to make large much larger chips than the competition and so we expect that there will be a shift towards those other two verticals. nd then in foundry services it means fabless chip designers which are a lot more than you might think especially because there are quite a few players that primarily design and develop IP and they have a little fab team to do R&D chips of like a couple qubits, but that’s something else than having the ability to build chips of hundreds, thousands, and then hundreds of thousands of qubits so that segment is where those players will come in. And the third one are obviously those players that need to have the capability to make chiplets or relevant sizes so they’re fairly sizable players.
Yuval: And if the company is very successful and I try to make the classical computing analogy you’ll be the TSMC of quantum or the Intel of quantum you know making two qubits and disregarding the current problems at Intel, right?
Matthijs: Yeah well that’s a very good question and also referring to the problems at Intel when we started we didn’t have those other two verticals we only had the fully packaged chips and so we also refer to ourselves as the Intel of quantum we don’t do that anymore for obvious reasons but it’s like you say you know these three verticals they basically correspond to three different business models right the fully packaged chips is Intel the middle one is TSMC and the right one is something like ASE or one of those large semiconductor packaging companies that most people have never heard of that are actually really large companies. And to some extent if you pulled it out 20 years from now you’re going to need to make a choice because if you’re doing say foundry service or fully packaged chips you’re competing with your own customers to some extent.
How that will pan out like in the current safety market I think it’s fair to have all three and how that will pan out yeah it depends on how the markets will pan out and to give that I mean we’ve seen that in the past as well right when we started we were the first to introduce commercially available quantum processors Soprano was the world’s first quantum processor that was commercially available. That has changed a lot of business models in the industry so there are now pure play system integrators parties that build and sell quantum computers that don’t make their own chips we’ve seen a massive acceleration in the field thanks to that we provided the chip for the first quantum computer in I think now five different nations including fairly large ones so that has really changed and then their response you could see players like Rigetti then rolleded out the Novera product. So you could see also other players reacting to our value proposition and you know and so you know the Rigetti sells computes right they want to sell quantum as a service and now they’re also selling chips so they are in that sense similarly positioned that in the long run you’re going to need to make a decision because you know otherwise you’re competing with yourself in that case and so or with your own customers and so in this phase of the market I think it’s fair to have all three but at some point so you know going back all the way back to your question right now we’re all of the above.
Yuval: I was listening to the Google Willow presentation a few weeks ago and they described some clever ways in how they’re overcoming you know two photon states and crosstalk and cosmic rays and resets and many other phenomena. Do you have to solve the same issues or does the architecture provide well maybe a completely different set of problems?
Matthijs: Yeah some of those yes, some of those no, and some of those we don’t know yet. For two-qubit gate protocols that’s much more on the design IP side or error protected qubits that’s not something QuantWare will do soon since we need to have design IP to ship our fully packaged products. We do have a sizable design team but we use generic transforms for that so that’s not really where our focus is and again that makes sense because let’s say for foundry services we’d want to service error protected qubits designers so those are not the type of problems we’re likely to solve. Things that are more related to the shielding the packaging, the routing of signals, through-silicon vias, deep stacking, that kind of stuff is really where our bread and butter is. And of course fabrication, so if we’re talking about coherence improvements, making very good TSVs, and doing clever things with housings, that’s really where our bread and butter is. And so in that sense when you refer to the Willow announcements I think that’s a really good example of why VIO is so important. It’s a fantastic improvement and it’s a big milestone and it puts the ball back in the court of the other platforms because it’s finally a big step is made again. But at the same time it’s a doubling of qubits in five years. That’s not fast enough. We’re never getting where we need to. We need to get that way up, and it’s hard if you don’t have a good architecture
Yuval: Let me try to maybe stretch the classical compute analogy maybe a little bit too far you’ll tell me but once upon a time CPUs had the math co-processors to the side and then over time they became integrated and then even CPUs have sort of some kind of GPUs integrated in them so it’s not just a set of chips so when we speak about error correction, error correction has a lot of classical components for the decoding and analysis and so on do you expect to integrate classical components or classical compute on your chip or is error correction completely external? I mean one could make a claim maybe not true that a million qubits is a fantastic achievement but it’s useless if you don’t have error correction attached to it.
Matthijs: So I fully agree. We need error correction. Superconducting qubits are great because they’re super fast and at scale speed matters because if you need to do a lot of operations, if you have a lot of circuit depth and your operations are a thousand to a million times slower then yes you could calculate it but it would take 10 years and therefore it wouldn’t be economical. So speed matters at scale but as we all know they’re noisy so you will need to do error correction and it better be fast because otherwise there goes your speed advantage and so the latency of that hardware and all of that is super important. So there we’re fully aligned. Then to your question of will there be classical computes on your chip, well not on the chip. What will be there in VIO itself at some point is SFQ which is classical compute, superconducting, but classical compute but that will be at various stages of the fridge. The reason for that is that classical compute generates heat which we don’t want or like. But if you talk about a million qubits in a QPU, you might have seen the paper, I think it was a collaboration between Rigetti and Riverlane, where they show real-time error correction for superconducting qubits using external hardware. There are a few of innovations like that make a system of that scale possible with what is actually right now very mature technology. It’s a very large size and there are if you start designing a system you run into silly things like the ports on that equipment are too large so that at that scale the physical size of those ports will be too large. You’ll run into issues like that but you don’t need to integrate that hardware on chip in order to go to a million qubits. You’d want to do that beyond that, but to get to a million qubits you don’t need to do it.
Yuval: How about connectivity? I know that in some other superconducting manufacturers they went from each qubit is connected to two or three and maybe soon six and something like that. Does that matter in your architecture do you have an improvement along those dimensions?
Matthijs: Well that’s really more of a design choice rather than a platform choice so again what VIO provides is a way to get the signal in from the top and then there’s a way to route that signal in various ways and so the default easiest or let’s say simplest way is to indeed have a 2D grid to do surface codes but if you want to make long distance connections then yeah then that’s totally possible and that’s really more on the design side we just provide let’s say the platform for people to design novel error correcting code that you know might use long distance copying.
Yuval: You mentioned VIO multiple times. Does that is that an acronym for something?
Matthijs: No it’s not. It’s just VIO.
Yuval: Very good. As we get closer to the end of our conversation, I’m curious, what have you learned about the market in the last 12 months that you didn’t know before that?
Matthijs: I definitely learned something about the power of marketing given what the stock market did after the Willow announcement which contains no very new, not a lot of new information compared to the arXiv publication that was out months before, so I think that’s definitely still both an opportunity and also a risk that this industry is still very hype-sensitive. Also learned a lot around you know we talked a lot about VIO whereas our website tells very little about VIO and so as a CEO it’s a clear indication that our storytelling needs to improve, a lot. And I think for the rest, I think for a lot of quantum companies apart from the 10x push at the end of the year in the stock market for a lot of quantum companies at least the ones that I’ve spoken to, it’s not been the easiest year. I think a lot of people feel they put in a lot of work and sold less than they wanted to, but then at the end of the year a lot of very interesting and cool stuff happened again. So in that sense I think it was a very interesting year. I am super optimistic for the next two years ahead. I think a lot more will happen than people expect. Yeah, I’m excited for what’s to come.
Yuval: And last, a hypothetical: If you could have dinner with one of the quantum or semiconductor greats, dead or alive, who would that be?
Matthijs: I think two names come to mind. One which is you know an obvious answer but for his strategic vision and how far he told ahead of what happens now, Jensen Huang. I think I think that’s such a good example of a company that was able to position itself in adjacent markets before getting the big prize. So that’s super-inspiring. And the other, who’s a personal hero of mine would be Scott Aaronson, who I admire a lot and who I think is one of the greatest thinkers in where the applications of quantum computers will be.
Yuval: Matthijs, thank you so much for joining me today.
Matthijs: Likewise, thank you.