Michelle Simmons, CEO of Silicon Quantum Computing, is interviewed by Yuval Boger. Michelle’s company uses phosphorus atoms in silicon-28 to create high-fidelity, low-noise qubits with exceptional coherence. She highlights their multi-nucleus spin registers for all-to-all connectivity and native multi-qubit operations, achieving what she describes as industry-leading fidelities. We discuss their three products—quantum machine learning, analog simulation, and fault-tolerant quantum computing—and their progress. We also discuss the expected date of their error-corrected system, the key to accelerating their roadmap, and much more.
Transcript
Yuval Boger: Hello, Michelle. Thank you so much for joining me today.
Michelle Simmons: Pleasure, Yuval. Nice to talk to you.
Yuval: So who are you and what do you do?
Michelle: My name’s Michelle Simmons. I’m the CEO and founder of Silicon Quantum Computing, which is a quantum computing company here in Sydney, Australia.
Yuval: And how does silicon quantum computing work? Is that like other silicon qubits or is there something unique about what you’re developing?
Michelle: Yeah, I actually think the company we have is globally unique because we’ve decided from the very get-go to choose what we think is the right material system and scalability system to give us the highest quality qubits. And so to achieve that, we went for very small qubits, so atoms in silicon. We’re using phosphorous atoms in silicon and using both the nuclear and the electron spins of those, but putting it inside an isotopically pure silicon-28 material. So it’s like the perfect environment to host the qubits. So it gives you small, high-quality qubits, but in a system that’s being manufactured at scale.
Yuval: And qubits are individual atoms, so all perfectly identical, I assume?
Michelle: Yeah, so the idea is really to have something that is very small, so that gives you the very large energy separations of very high-quality systems. And atom qubits have demonstrated some of the longest coherence times of any solid-state system, so up to 35 seconds for the nuclear spins. And I guess the advantage of going very small is that at these small systems you do get these natural qubits. You don’t need lots of materials engineering to create them because they naturally form just by placing the atom inside the silicon crystal. But I think the thing that we got very excited about is in our quantum processing unit we actually only have two atoms in the whole device. So we have phosphorus atoms and silicon atoms and that simplicity really allows us to get this incredibly high quality but also very strong stability. So it’s a perfectly epitaxial system, fully crystalline, that means very low noise and because we’ve got that nice silicon environment which we can isotopically purify at very low magnetic noise and it’s that combination of low charge, low magnetic noise, very small high purity system that gives us the high fidelity that we’re looking for.
Yuval: If you have such small qubits, does that mean you need to have really small wires? How do you control them?
Michelle: So one of the things, this is a great question, is right from the outset we realize that we’re going to have to design all the circuitry as well. We figured out that within about the first 5 – 10 years that we can actually make the circuitry from phosphorus-doped wires as well. So we don’t need to use metal wires that come all the way down to the qubits, but we can use these very narrow phosphorus-doped wires. And then again, because they’re epitaxial, they’re very fast. We can get basically gigahertz frequencies down towards where our qubits are. We can make them all the way down to the size of the atom itself, get way down to the atomic scale. But again, they’re epitaxial so they’re low noise. So very low noise circuitry to attach to our qubits. And literally only phosphorous and silicon in the active region of where the qubits sit.
Yuval: I think about superconducting qubits, I know yours are different. One issue that comes up often is connectivity. How many qubits does each qubit interact with? What’s the answer in your system?
Michelle: Yeah, so we have these registers we call the multi-nucleus spin registers and that basically allows us to control the nuclear spins using the hyperfine with the electron spin and so we can actually get multi-nuclear spins in within one register with all-to-all connectivity and that’s what we’re actually using and have demonstrated recently to give us these very high fidelity algorithms and not just the individual gates. So we’ve got single qubit gates now at four nines fidelity, two- gates at 99.5 fidelity, readout 99.95%, but we’re now able to run algorithms with our system. So basically without doing error correction we can still get 98.87% on Grover’s algorithm with a four- system running a three- algorithm. And so this all-to-all connectivity of these nuclear spin registers is quite unique. It gives us something called a multi- gate, so we can do one operation on the electron spin and that automatically can entangle multiple nuclear spins. And because of the hyperfine interaction, that gives us the ability to individually address them without creating errors on neighboring qubits.
Yuval: So you can do a Toffoli gate natively, essentially.
Michelle: That’s right, that’s right. And I guess in order to get this to be high fidelity, we’ve really focused heavily on the precision manufacturing. So we have sub-nanometer precision, so we can bring our readout sensors very close, that allows them to be very, very fast. We can engineer those sensors to be very, very small. So there’s a very strong on-off ratio, so we get a very strong signal with very low noise. And that’s what’s giving us this kind of precision is giving us the speed and the fidelity, which is basically the quality of our qubits.
Yuval: – You mentioned four qubits per register. I’m guessing this is just an interim milestone, but do you expect a larger computer just to have more qubits in a register or multiple registers?
Michelle: – Both, both. So we’ve been assessing how many qubits we can get per register and so far we’ve got up to six. And the exciting thing about that is you can have a single control gate that addresses six qubits. And so literally in that system, with one signal you can individually address the qubits. Whereas typically in the solid state a lot of people have six gates per qubit, so six physical metal gates to create a qubit. Because we have a naturally forming qubit we just need the gate to address the qubit. So we can get six qubits per gate. And that is really fantastic for scaling. We’ve shown that we can actually now connect these registers together. So that’s also very exciting and we get very high fidelity operations between the registers.
Yuval: So my question on the connectivity, maybe I should have addressed it about the registers, not the qubits. So how many registers does each register connect to?
Michelle: So the registers will connect in a kind of surface code architecture. It all depends on the error correction code that you’re running. And so we’ve got multiple different architectures for different codes that we’re assessing at the moment for our system. Obviously, surface code is very popular, but there are other codes that we think our system actually may be even better to address, giving us lower overall error rates.
Yuval: – Sometimes in chip manufacturing, there are yield issues, right? So for instance, you mentioned four or six qubits per register, is that across all registers, or do you sort of make the chip and they say, Oh, I found that I only have five atoms here and…
Michelle: You’ve hit for us one of the key strengths of the company is that we manufacture with precision. And the whole point is that we literally engineer what we want and we figure out how to do that, both at speed and to do it reproducibly. And so speed for us is really very important. So we can literally design a chip, manufacture it and test it within a week. And that gives us a very fast cycle time, but right from the beginning, the whole reason why we went to this kind of atomic precision technique was so that we would make them reproducibly. Because there’s nothing worse, honestly. I mean, my experience in my early career was making lots of devices and trying to find one that works. It’s a soul-destroying process and I don’t believe you’ll be able to scale if you have that. So you have to be able to have reproducible manufacturing.
Yuval: If the cycle time is so quick, does that mean that you’re looking to build a general-purpose processor or do you think you’ll build algorithm-specific chips?
Michelle: So, again, one of the nice things about our system, because of that sub-nanometer precision, we can do both. And so within the company we actually do have what we call three products that we’re addressing.
One is in the machine learning space and that is to use essentially analog systems where we basically take classical data through what we call a quantum feature generator. It allows us to take that data to much higher dimensionality and then we can pass that back through to classical machine learning and it either increases the speed or the accuracy of classical machine learning. And that is a real product that we have available now. So we have customers using it, we’ve been testing it for the last two years with those customers and we see an advantage. So we’re very excited about that. We haven’t officially launched yet but we’ve got the results on the ground and with customers. So that would be our first product.
Our second product is in the analog simulation space where we’re literally looking at how to mimic certain molecules. This is an area that we think will explode, it’s early days for analog simulation, but we started off with something like polyacetylene. It’s a very small molecule with 10 different atoms in it. Something you can classically simulate and we showed that we could keep it quantum coherent over the size of the molecule. And just to step back on that, essentially what we’re doing is mimicking a carbon single and a carbon double bond. And in order to mimic that in real time you have to have that sub-nanometer precision to mimic it. And that is what we have. And so we’ve realized now that we can rapidly scale that to many many thousands of different atoms. And so there’s a lot of different chemicals and materials we’re looking at in that space.
And then our third product is obviously the full-scale error-corrected quantum computer. That’s the big game for us. That’s where most of the companies focus on. And really that’s all about taking those high quality qubits that we’ve already demonstrated. With the Grover’s algorithm, we’ve not even used error correction. So the fidelity we’re getting is kind of a native raw fidelity. And so now we’re implementing larger systems and seeing what happens as we add more and more qubits. Is that fidelity maintained?
Yuval: On the second product, is the output basically the ground state of that molecule or is it something else?
Michelle: No, it’s the ground state of the molecule. That’s what we’re looking at at the moment. So on the polyacetylene, we were just looking literally at passing current through and then predicting what the current through the molecule would be. And then actually measuring it directly and showing that it matched what the theoretical prediction was. But yeah, on some of these more complex devices, it will be ground states.
Yuval: Sometimes when manufacturers talk about having very high quality qubits and very high quality gates, they sort of sidestep the error correction question. So they kind of say, oh, we don’t need error correction because we’ve got such wonderful qubits. But that’s not what you’re saying, right? You still need to have error correction.
Michelle: Yeah, yeah, absolutely. this is also something that I do find frustrating is that the details matter. And so we typically, when we quote our fidelities, we include all the losses. So the numbers you’re getting is with losses included. A lot of people will not mention losses or not mention their SPAM errors, their state preparation and measurement errors. We include those in all the fidelities that we quote. And so I guess what’s remarkable for us is those numbers are incredibly high, including the errors. And so we know that as we add more error correction into the system, which we’re starting to do, that that’s going to either get better or it’s going to be maintained. And the key thing is to show that the algorithmic fidelity remains high, not just the individual qubits. You’ve got to actually run algorithms and show that all the way through that you’re maintaining the high quality coherence times and you’re able to operate quickly whilst keeping those gate fidelities high.
Yuval: When we look at error correction and look at decoding, that usually requires classical hardware. So the classical hardware would be external to the chip or is it somehow embedded on it?
Michelle: At the beginning, I thought I was really only going to be making the QPU. We really rapidly realized that a lot of the control electronics that was available was just not suitable. So we’ve been designing control chips that sit near our qubit chips. We’ve also been designing the FPGA controllers. So as time has evolved, we’ve actually now built through the full stack and we’ve made our processes available through the cloud. So the customers that we have are actually using it through the cloud at the moment. And that basically requires that you, by having that full integration in-house, you can see how, as you add every layer to the previous layer, you get the best out of the layer below. So whether it’s the control chip next to the chip, whether it’s the high frequency electronics, whether it’s the software embedded software, or all the way through to the algorithm development, we get the whole thing in one go.
Yuval: So just between us, when, when is the product launching? I mean, when could, uh, everyone use this?
Michelle: So I think we’ll be launching it sometime this year. It won’t be at the end. It will be like mid-year.
Yuval: Very good. Tell me a bit about the company. How large, how are you funded, are all the offices, is the only office in Australia? What should people know about the company?
Michelle: Yeah, so the company really evolved out of a Center of Excellence here in Australia. So Australia has these large centers of excellence which have about 200 people. I’ve been the director of that for the last 15 years, but it started in 2000. Straight away, I guess one of the key things is, in 2000, there was a group of us that decided we’re not going to adapt existing technologies to try and make a quantum computer. So a lot of people were using, for example, gallium arsenide. I was at Cavendish in the UK making very high quality gallium arsenide chips. And I was realizing then the problem of manufacturability. Kind of the top-down metal above some kind of heterostructure was a very difficult way to reproducibly make the same thing. And the material is not great because both gallium and arsenic have spins in them. And so I was looking at what’s the best material to choose. And a guy called Bruce Kane had written this paper in Nature in 1998. He was from Bell Labs originally and I knew him from Bell Labs because I was at the Cavendish and we were really competing to make the highest quality transistors at that time and looking for very pure quantum effects. So the higher the quality of the system you have, the better the quantum effects. And then how do you make it reproducible to understand it? And so he came out with, if you want to make a quantum computer, his thesis was you would make it with atoms and silicon. And for me, coming from a very strong materials background, I realized that the reproducibility of atoms is far better than trying to squeeze something from a traditional top-down technique, where you’re taking it to the very limit to capture a single electron. So that was the kind of thesis at the beginning. Let’s choose a material that has high quality from the get-go but that is also manufacturable, which is why the silicon was so critical.
Anyway, so I came down to Australia in 1999 and we set up the first Center of Excellence, which is now still running (It’s in its third round) and that was focused around can you actually put a single atom in place? What’s the manufacturing challenge of doing that? If you do it, do you actually get the high quality you expect? So we literally pioneered the whole technique of how you put a single atom in, how you bring the leads down to that device and then how do you control it? I think we are the only company, I’m pretty sure, the only company in the world that manufactures with atomic precision. And it’s really been understanding the materials to give us the high quality. Then we started finding that the qubits were very high quality. We had companies in Australia wanting to work with us, so Commonwealth Bank and Telstra, recognizing we had high quality systems. And so then in 2018, we set up the company with five shareholders, the Commonwealth Government, Commonwealth Bank of Australia, Telstra, State Government and the University of New South Wales. So it was quite an unusual set up. It took probably about three years for me to negotiate that set up with those parties to get it underway. And I guess just to put that in context for the Australian system, that’s, it’s not a standard thing. There is no easy route to create a company. So this was me creating the structure that I needed to create a company in the first place. And like I said, at the beginning, we were focusing on just a hardware chip. And as time has evolved, we actually realised we have all the expertise within Australia to build all the different layers of the stack. And so now we have about 70 engineers in the company. The company has been going roughly about seven or eight years now. And like I said, we’ve got, there’s three products that are coming out. Two of them are available on the cloud at the moment. And the third one we’re working on scaling and getting that on the cloud in the future.
Yuval: You know, in job interviews, sometimes people ask, you know, what is your weakness or what do you think is the weakness of this approach? What are the things that you say, “Oh yeah, I wish I had the X of, you know, neutral atoms or something?”
Michelle: So honestly I don’t think there is a weakness and that’s from having worked in the field from the beginning we set out to de-risk all the way along the way. We’ve always had a 10-year horizon view about what are the challenges we’ve got to face and at each stage, you know, if you’ve never done something before, you never know it’s going to work. But we had a pretty high confidence level that those things would work. And I would say that consistently over the last 25 years, we have found things that are way better than we expected. So the low noise in the system, the addressability of the qubits, the fact that we can run algorithms without error correction with such high fidelity, I wasn’t expecting that, to be honest. So the thing that probably is the hardest thing is that we are an Australian company. And so in the global world when you’re dealing with companies like IBM and Google and Microsoft, they’re so big, they’re so well known, that the perception is can a small company in a country like Australia compete? And I guess my instinct on that is actually that Australia has many competitive advantages. It’s a very collaborative nation, so it works well in great big teams, but it’s also highly competitive. Anyone that watches sport will know that. And so there’s a very unique culture down here that I think in addition to the technology will give us advantages.
Yuval: How did you feel when the Australian government decided to give a billion Australian dollars to a company that’s not yours?
Michelle: In some ways it’s very extraordinary, but the Australian government is very keen to grow the ecosystem down here.
Yuval: Does the system need cryogenic cooling, by the way?
Michelle: Yes, yes. So at the moment in the solid state systems, in general, fidelities are always better the lower your temperature. And so for that reason, we do believe we’re going to be using dilution fridges all the way through. We have got results, and there will be some coming out soon, that show that it operates all the way up to 4 Kelvin. It operates pretty well up to 4 Kelvin, but the fidelity drops by about half a percent. So it’s something small, but in the game where fidelities matter, at the moment we’re maintaining a dilution fridge approach, as well as investigating how well we can get it working at high temperatures.
Yuval: And if you think about, say, a computer with 10,000 qubits, does it all fit on a single chip or do you have to network multiple chips?
Michelle: Yeah, so the great thing about what we’ve got is that we get huge numbers of qubits, thousands of qubits per QPU, but they will all fit with our control electronics onto a wafer that fits inside one dilution fridge. And so that, in terms of connectivity, the connectivity is just on-chip connectivity that you’d have with any kind of device in silicon, but it fits in one dilution fridge and I think that’s a massive advantage for us. We don’t have to develop new technologies that we don’t know how they’re going to behave. Everything we do is with existing technology and materials.
Yuval: You’ve been doing this for many years and it sounds like the company is, I think, what, seven or eight years old. What have you learned over the last year that you didn’t know about quantum computing or about your particular technology?
Michelle: In the last year, I think, I mean, honestly, our approach has been rigorous and systematic, and so we cover all bases at all times. So for our approach, I think we’re kind of locked into all the challenges we have and just marching through them. I was surprised at the cold atom technology coming through so quickly, I thought that was great. I see a movement across a lot of different companies globally. And I think each one of them has a challenge, a kind of, for me, an insurmountable challenge that they would have to deal with. The way I look at our company is that the number of challenges that we have are smaller and more surmountable. One other kind of reflection is that I know that details really matter. And I think one of the things I find frustrating, I think a lot of us in the field do, we have a motto in the company, keep it real. A lot of people are presenting results where they’re kind of not putting the full information out there, so they’re not accounting for losses or, you know, the press release doesn’t match what the paper says. There’s a lot of kind of strangeness that’s happening in the field, which, which makes it, I think, very difficult. And so, yeah, keep it real, keep it systematic, be thorough, details matter.
Yuval: In recent weeks, a lot of big company CEOs have decided to weigh in on their prediction of when quantum computers will be useful. Do you want to contribute to that discussion?
Michelle: Uh, so, well, so for me personally, I already have something that is useful. So we have our first product, which we know is useful because customers are telling us. So from that perspective, for me, the customer validation is the most important part. Um, you know, it’s very difficult to say how good your system is against one another. Customer telling you they like it and they want to buy it. That’s good validation. For the longer term, we’ve kind of mapped out when the full-scale error-corrected system will come because that’s obviously the big one. And for us, we’ve really just taken a very, again, pragmatic approach. We’ve looked at what the silicon industry has done with classical computing. The first transistor, 1947, the first integrated circuit, 1958. The first kind of commercial product for about five or six years after that, 1971 for the first industrial-scale computer. And we literally have mapped our results to that. So our first single item transistor was 2012. Our first integrated circuit was 2021. So it was nine years, not 11 years. Our first product we were predicting would be 2028. We think we already have it this year. Arguably we had it last year, but we’ve been testing it, but we’ll release it this year. And then 2033 would be the date when we think the error corrected system will come. And now that’s obviously based on literally just mimicking what the industry has done in the past. So one might hope that we would go faster knowing how it’s worked previously. And so, yeah, 2033 is something I’m comfortable to say that we’ll be delivering by.
Yuval: If I were a board member at your company, I’d say, okay, well, what do you need to accelerate that? Why not 2031? What do you need?
Michelle: For us, literally everyone would tell you this, but it’s dollars. So we basically, we have a manufacturing plant down here. You’ve already seen how fast our chip cycle times are in a week. So we could at this point, given the results that we’re getting now, start to double the size of our operation. And that is exactly what we’re planning to do. And once you double the size of your operation, I think it will accelerate the roadmap that we already have.
Yuval: As we get closer to the end of our conversation, you mentioned customers. What type of customers are these? Are these HPC centers? Are these enterprise customers? Are these researchers for academia?
Michelle: No, no, they’re companies that are buying our system and they go across from transport systems, logistics systems.they’re going across pretty much every industry sector. So if you can demonstrate that you have a processor that enhances the accuracy of machine learning, particularly if you have sparse data sets or you have time series data, there are a lot of people that are very interested in that problem. And so we have three customers already tested. We’ve got another one that’s just started and we’ve got about five waiting to get in. So yeah, it’s starting to grow quite rapidly and it’s honestly very exciting.
Yuval: So last, a hypothetical: if you could have dinner with one of the quantum greats, dead or alive, who would that person be?
Michelle: To be honest, I’ve got heroes that are unfortunately dead, and they go all the way back to Michael Faraday. Actually, I wouldn’t have dinner with just one, I would have a little mini group of them. I would have Michael Faraday. He was a self-taught experimentalist, very much somebody who had a very good instinct. He tried experiments out over and over again and then got the results and came up with electromagnetism. He was a chemist by background. I’ve heard that Einstein had a picture of him on his desk, so he valued his instinct and ability to iterate. I think John Bardeen is a theorist that’s obviously won very exceptionally two Nobel Prizes in physics, one for the transistor and one for superconductivity. The two areas that I’m actively looking at the future of those two areas. So having him as a theorist in the room would be great. And also Gordon Moore. Gordon Moore because he was , again, he was actually a chemist by training. Then trained in the field, so he went to industry at Fairchild, but he was someone that transitioned across different boundaries from academia to industry. But he had that relentless pursuit of delivering an outcome, iterating and understanding and strategically planning what needs to be done next to get the best next thing. And so having the combination of those three together, I think would be a very interesting conversation based on, you know, understanding what’s important, iterating and where it needs to go. It would be fascinating conversation. I’d love to have that dinner with those three.
Yuval: Wonderful. Well, Michelle, thank you so much for spending some time with me.
Michelle: It was a pleasure. Very nice to meet you.