Beyond the Qubit

Stefan van Waasen on cryo-CMOS scaling, multiplexing, and the 10k qubit milestone

44 min Listen to the original Episode page
Quantum ComputingCryo-CMOSSemiconductorsHardware ArchitectureRISC-VSupply ChainMemory TechnologiesScaling

Summary

Stefan van Waasen, co-founder of IceCirc, discusses the cabling bottleneck in scaling quantum computers and how cryo-CMOS architectures can mitigate it. He argues for keeping high-frequency signal generation at room temperature while moving signal shaping and biasing circuits to the 4 Kelvin stage to reduce heat load and cabling complexity. Van Waasen highlights RISC-V processors as ideal local controllers due to low latency and open-source licensing, while noting that SRAM is not currently a scaling bottleneck for silicon spin qubits. He identifies reaching 10,000 qubits as the key milestone indicating that control electronics problems are solved, with market acceptance being the primary current constraint.

Key takeaways

  1. 1
    Multiplexers are predicted to play a much more important role in the future because they reduce the number of lines from room temperature into the cryostat, addressing the scaling bottleneck.

    Unidentified speaker (speaker_1) · 2:40

Episode sections

  1. 0:13 – 1:14

    Introduction and Guest Background

    Host introduces Stefan van Waasen, co-founder of IceCirc and professor at the Jülich Centre for Neutron Science. Discussion focuses on scaling quantum computers and the cabling bottleneck that cryo-CMOS aims to solve.

  2. 1:14 – 3:06

    Control Electronics Building Blocks

    Van Waasen identifies AWGs, DACs, and multiplexers as main building blocks. Notes multiplexers will become increasingly important for scaling.

  3. 3:06 – 5:10

    DACs and RF Control Mechanics

    Explanation of DACs converting digital signals to analog voltages/currents. RF control involves shaping high-frequency signals using envelopes generated by DACs for precise qubit pulses.

  4. 5:10 – 9:07

    Multiplexing Benefits and Trade-offs

    Multiplexers reduce cables entering the cryostat. Benefits include reduced hardware, lower power, and cost savings, but only if overhead is reduced.

  5. 9:07 – 15:25

    RF Control and Cryogenic Partitioning

    Van Waasen argues high-frequency signal generation should remain at room temperature due to power consumption. Cryo-CMOS should handle signal shaping at 4 Kelvin, as shown by Intel and TU Delft.

  6. 15:25 – 17:35

    RISC-V Processors for Local Control

    RISC-V processors are discussed as local controllers at cryogenic levels due to low latency and avoidance of ARM royalties.

  7. 17:35 – 30:03

    Heat Management and Component Placement

    Partitioning strategy: biasing circuits for semiconductor qubits move to lowest temperature stages to reduce cabling (12 to 2 cables), while RISC-V and power supplies stay at 4 Kelvin to avoid heat load at milli-Kelvin levels.

  8. 30:03 – 34:22

    Memory Technologies and Scaling Challenges

    Need for local, distributed memory to overcome latency and power bottlenecks. SRAM is current standard, but DRAM could be beneficial at cryogenic temperatures due to reduced leakage. Cost and scalability beyond 10,000 qubits remain open questions.

  9. 34:22 – 35:42

    Local Square Memory Architecture

    Modular architecture where each unit cell contains local square memory, enabling copy-and-paste scaling without external memory access.

  10. 35:42 – 37:37

    In-Memory Compute and Neuromorphic Computing

    Potential for in-memory compute (neuromorphic computing) in quantum systems. Currently too distant for immediate implementation in cryo-CMOS designs.

  11. 37:37 – 39:21

    SRAM Scaling Bottlenecks and Cost

    SRAM is not currently the limiting factor in form factor, size, or cost for silicon spin qubits compared to other unit cell functionalities.

  12. 39:21 – 40:27

    Market Acceptance and Foundry Availability

    Primary constraint is market acceptance and lack of system-level examples. ST offers FDSOI in Dresden, but reliance on global foundries poses future constraints.

  13. 40:27 – 41:07

    Technology Roadmap and Feature Sizes

    Cryo-CMOS is far from hitting physical limitations like feature size scaling. Monitoring when shifts to smaller feature sizes or different technologies are necessary.

  14. 41:07 – 43:41

    Early Stage Market and Scientific Backbone

    Company founded early with no prior market interest. Emphasis on maintaining scientific backbone separate from operations to anticipate tech evolution while establishing market.

Claims worth verifying

Stefan van Waasen is co-founder of IceCirc and a professor at the Jülich Centre for Neutron Science.

Unidentified speaker (speaker_0) ·0:22

How to verify: Check IceCirc's official website for team bios and university faculty listings for Jülich Research Centre. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Current main building blocks for quantum control are AWGs from Zurich Instruments, Quantum Machines, and Qblox.

Unidentified speaker (speaker_1) ·1:44

How to verify: Check product catalogs of Zurich Instruments, Quantum Machines, and Qblox for AWG offerings. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Multiplexers allow one signal line to be distributed to multiple outputs, reducing cable count.

Unidentified speaker (speaker_1) ·2:40

How to verify: Review technical datasheets for quantum multiplexers from companies like Quantum Machines or Qblox. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Intel and TU Delft demonstrated cryo-CMOS frequency generation at the 4 Kelvin level.

Unidentified speaker (speaker_1) ·9:44

How to verify: Search for joint publications or press releases from Intel and TU Delft regarding cryogenic CMOS and quantum control. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Moving biasing circuits to the lowest Kelvin level reduced cables from 12 to 2 for semiconductor spin qubits.

Unidentified speaker (speaker_1) ·18:47

How to verify: Look for research papers from IceCirc or Jülich Centre demonstrating this specific cable reduction experiment. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

RISC-V is chosen to avoid royalties associated with ARM.

Unidentified speaker (speaker_1) ·16:35

How to verify: Check RISC-V International licensing terms versus ARM's licensing model. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Primary source

Scaling beyond 10,000 qubits presents significant cost issues.

Unidentified speaker (speaker_1) ·32:21

How to verify: Review industry reports on quantum hardware scaling costs from firms like McKinsey, BCG, or quantum-specific analysts. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

DRAM leakage is lower at cryogenic temperatures than at room temperature.

Unidentified speaker (speaker_1) ·31:05

How to verify: Consult semiconductor physics literature or memory manufacturer technical notes on cryogenic DRAM performance. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Primary source

Superconducting qubits are sensitive to temperature changes, requiring flux control at the 4 Kelvin stage.

Unidentified speaker (speaker_1) ·21:16

How to verify: Review technical specifications for superconducting qubit control electronics from companies like Rigetti or IBM. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

IceCirc is preparing a publication with partners including Arc Systems on cryogenic memory.

Unidentified speaker (speaker_1) ·31:52

How to verify: Monitor academic pre-print servers (arXiv) or IceCirc's news section for the mentioned publication. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Local square memory allows each unit cell to have its own memory, enabling copy-and-paste scaling without external memory access.

Unidentified speaker (speaker_1) ·34:22

How to verify: Technical whitepapers or patents from IceCirc describing their modular architecture and memory integration. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Neuromorphic computing and in-memory compute are being researched by the speaker's institute but are far from implementation in cryo-CMOS.

Unidentified speaker (speaker_1) ·35:43

How to verify: Research publications from Stefan van Waasen's institute on neuromorphic computing and cryogenic CMOS. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

SRAM is not currently a scaling bottleneck for silicon spin qubit quantum computers in terms of form factor, size, or cost.

Unidentified speaker (speaker_1) ·37:50

How to verify: Technical assessments or papers from IceCirc or partner quantum hardware developers on unit cell resource allocation. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

ST offers FDSOI technology produced in Dresden.

Unidentified speaker (speaker_1) ·39:59

How to verify: STMicroelectronics product documentation or investor presentations confirming FDSOI production in Dresden. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

The speaker's company is far from hitting physical limitations like feature size scaling in cryo-CMOS.

Unidentified speaker (speaker_1) ·40:27

How to verify: Roadmap documents or technical briefings from IceCirc regarding their process node and scaling plans. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Scaling beyond 10,000 qubits in solid-state quantum computers requires cryo-CMOS.

Unidentified speaker (speaker_1) ·43:18

How to verify: Technical arguments presented in IceCirc marketing materials or industry analyses on quantum control wiring challenges. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Do your own research

  • Verify the technical claims regarding cable reduction (12 to 2) by reviewing Intel and TU Delft publications on cryo-CMOS demonstrations.
  • Check the current market share and product offerings of Zurich Instruments, Quantum Machines, and Qblox for AWGs and DACs.
  • Assess the supply chain risks for STMicroelectronics' FDSOI technology in Dresden and identify alternative foundries for cryo-CMOS production.
  • Review academic and industry literature on the thermal performance of RISC-V processors at 4 Kelvin versus milli-Kelvin temperatures.
  • Investigate the cost and scalability challenges of SRAM versus DRAM for quantum control systems beyond 10,000 qubits.
  • Look for independent validation of the claim that multiplexing reduces overhead only when it provides clear system-level benefits.
  • Research the current state of neuromorphic computing integration in quantum systems to assess the timeline for in-memory compute.
  • Verify IceCirc's current technology stack and whether they utilize ST's FDSOI or other foundry technologies.
Full transcriptexpand

0:13Unidentified speaker (speaker_0): Welcome to Beyond the Cubit. My guest of today is Stefan van Waasen, and I'm quite excited that Stefan is

0:21Unidentified speaker (UNKNOWN): here.

0:22Unidentified speaker (speaker_0): Stefan is co-founder of iCirk,

0:24Unidentified speaker (UNKNOWN): and

0:25Unidentified speaker (speaker_0): he is professor at the Julie Forensig Centrum. And what I think is really exciting is that if we want to scale the amount of qubits, the cabling could be a potential bottleneck. And Craio Simos is able to solve or contribute to to to solving that that bottleneck. And Stefan, he he's worked for more than a decade on cryosemos on cryoelectronics and system architecture. So I'm keen to learn more. Stefan, one of the things, if I understand it correct, is by C cryosimos is that you bring more control electronics closer to the chip. Is that correct?

1:14Unidentified speaker (speaker_1): That's absolutely correct, yes.

1:17Unidentified speaker (speaker_0): Okay. For the listeners that are not familiar with the control electronics, can we maybe describe the main building blocks? And I have three building blocks, but correct me if I'm wrong. I see DUX DAC, I see multiplexers and r uh r radio frequency generation and control. Are these the most important building blocks or am I missing others?

1:44Unidentified speaker (speaker_1): Yes, in principle you are right. Uh from today's perspective is a little bit different. That's a little bit already in the future, sought into the future. Today the m main building blocks are AWGs, WGs yeah, here from from from from uh companies dil uh like like Zurich Instrument, Quantum. Machines, Q blocks or whatever. So where you generate these control signals with, they are also generating this this uh RF control signals and so on. But but from a from a general general point of view that are mainly ducks, DUCS, so d digital to analog converters. And as I said, they are also generating generating the RF signals. Multiplexers are not used so much today, but into the future they will play a much more much more important role. Yeah, because multiplexers gives you the possibility to bring less lines from the top, so from room temperature into the cryostat and

2:48distribute or move things. signal distribution to the uh

2:53Unidentified speaker (UNKNOWN): to

2:53Unidentified speaker (speaker_1): the bottom of the cryostat to the coldest stage or to the colder stages.

2:58Unidentified speaker (speaker_0): Okay. Can you explain the basics what exactly is a DAC?

3:06Unidentified speaker (speaker_1): So what you what you normally do if you if you do signal processing or if you would like to get something under control, you generate digital signals. So you generate the huge number of zero and ones. And then out of this you you know what signal height you, for example, of or what signal frequency, or in the time domain to be honest, how a signal looks. If you have a sinusoidal signal, if you have a rectangular signal. And that is what what a DAC, so a digital to analog converter, is converting into the analog domain again. So to make out of this numbers real voltages or currents again. That is what a DAC is doing. So a digital to analog conversion. Nothing else. So generate out of numbers voltages. Nothing else.

3:55Unidentified speaker (speaker_0): Okay. And and XI I control,

3:58Unidentified speaker (UNKNOWN): the

3:58Unidentified speaker (speaker_0): high of the frequency, and the set control, are they related to DACs?

4:03Unidentified speaker (speaker_1): Yes,

4:03Unidentified speaker (UNKNOWN): they are related to

4:05Unidentified speaker (speaker_0): different.

4:06Unidentified speaker (speaker_1): No, no, they are related. You also need DACs to get RF control, for example. Because what you are doing in RF control is nothing else. You have a certain frequency, so a certain sinusoidal signal, which you would like to generate. And this needs to be, for example, but you do not need to have it continuous, you need something like high frequency. pulses, so which is m which means you are just shaping this frequency. So you are ramping up the amplitude, then you keep it constant, and then you ramp it down again. For this you also need DACs. So that's called uh that's uh that's called the the so or that's a so-called so called envelope. You just put an envelope with an DAC on such a high frequency thing. signal and out of this you make in high frequency pulse for it for finally yeah.

4:56Unidentified speaker (speaker_0): Okay. Thank you for explaining. And the second building block multiplexers. What are they? What exactly do they and why are they important?

5:10Unidentified speaker (speaker_1): An multiplex a multiplexer you you generate for uh let's make an easy example. You have a one to four multiplexer, which means you get in one signal, which you like to be uh dynamically distribute to different outputs. Yeah, because you would like to control it to different qubits, for example. So what you do is you so you bring one signal in, which means finally one cable, and at the on the outside, so at the output of this multiple. multiplexer, you have, for example, four cables. And then you decide inside the multiplexer by programming it to which output one of this or this one input is going. But you can have it in different variants. You can have one to eight, one to sixteen, th four to eight, for example. So that's the combination is a little bit uh it's it's uh let's say what is required, but finally the target is just to make one uh just distribute signals. It's like a small distribution network. Yeah, how how to distribute one input signal to to different outputs, yeah? Whatever is requested where you need it, what you need, and so

6:22Unidentified speaker (UNKNOWN): on. And

6:23Unidentified speaker (speaker_0): what are the pros and cons of of multiplying? multiplexers? For example, I was at EAC, Ecoc, uh last week, and what I was wondering is is okay, the there was Photon Bridge, who said okay, this is for AI, this is not for quantum. But they said okay, with with multiplexing um we got significant benefits. But what are the disadvantages of multiplexing? Well, at at what cost ca it brings a lot of benefits. But i is it a zero sum game or or are there also disadvantages?

6:59Unidentified speaker (speaker_1): Yeah. Multiplexing in optical is a little bit different. So we are talk talking about different types of multiplexing, finally. But it in i but in on the electronics electronic side is is clearly a benefit because in uh especially uh sp if you're controlling and operating qubits, you do not need all signals always at the same time at every qubit, yeah. So and so before using cables which are unused coming from room temperature, you just do it with multiplexing, which means you can you can save some hardware on the room temperature side, for example. So you do not need you you need instead of four uh signal generators, you need only one signal generator, but neverthel nevertheless can serve four qubits, yeah. So if this is not possible to to multiplex this, which means finally nothing else or not everything is needed at the same time, then you cannot do it ev ever anything else than just unst and instead of using one line, you need to use four lines. But there are many things where you can make use of it. So it's finally saving power, it's saving hardware, so it's reducing cost, and so on. So that's normally it's a huge benefit, and we only use

8:15multiplexers

8:17Unidentified speaker (UNKNOWN): if

8:17Unidentified speaker (speaker_1): it is useful. So if they are just adding overhead, we just let it out and say okay. And the advantage advanta another advantage is multiplexing almost is very low power consumption. The main power if if is if you change settings, which means if you would like to switch the signal from one one output to another, then you generate and you you generate a little bit losses, which is unfortunate, but even instead of force four signals, for example, it's always beneficial from the overall power budget but it is normally always beneficial.

8:53Unidentified speaker (speaker_0): Okay. Um

8:54Unidentified speaker (UNKNOWN): and

8:55Unidentified speaker (speaker_0): the last but not the least, the radio frequency control.

8:59Unidentified speaker (UNKNOWN): Can

8:59Unidentified speaker (speaker_0): you explain the basics? Why why does it matter for the qubit?

9:06Unidentified speaker (UNKNOWN): Okay.

9:07Unidentified speaker (speaker_1): There was yeah, there Th was uh radio frequency control. Is that th that's a that's a that's a very important point. And there we are uh to my opinion uh still at an early stage. So so you need this you need finally this radio frequency control to make major operations on the qubits. Yeah. Which means you are in very high frequency range. High frequency always means huge power consumption. So you most likely uh will generate these frequencies or this basic frequencies at room temperature because their power consumption is not so crucial. And what you do then on the cryo part is that is for example what what Intel together with Theo Delft showed in a very excellent way quite some years ago already. To use one frequency and generate let's say sub frequencies out of this for different channels works excellent. Which means cryosemos also can help you there to to avoid that you need to generate too much frequencies, so which means again cost it's a cost issue again,

10:17and distribute it to different channels. So frankl f finally this RF control is a very nice mix of a little bit different way of multiplexing you can use mit you can do with cryosemos. And to to buy this to solve also the issue to minimize power consumption, what you need to generate or need to bring to the uh

10:45Unidentified speaker (UNKNOWN): to

10:45Unidentified speaker (speaker_1): the qubit level or to the coldest level where it generates then a problem of heating it up. Yeah. So IF control is in a certain way in a certain way a subtopic of this duck and multiplexing things in a certain way. Because for these for these IF control in the fridge, ducks are very a very crucial point, for example.

11:09Unidentified speaker (speaker_0): Okay. Stefan, thank you for for bringing everybody up to speed with from the the basics. basics. Now let's dive a little bit deeper. What tasks does the control Electronica? And which which tasks can stay at room temperature? Which tasks have to go to to to maybe a few millikalvin and which has to go maybe to 200 two hundred millikalvin? And maybe starting with the X i XI control for the higher frequency.

11:43Unidentified speaker (speaker_1): As I said before, the the generation of this high frequency should stay definitely at at room temperature. Because there's one physical there's one physical effect. Even if you would assume from a first guess bringing for example a high frequency control to cryo cryotemperatures temperatures would give you some advantage That's not the case, yeah. So generating high frequency signals at room temperature, there is no benefit to bring it in. By the way, I would turn around the question in a certain way what is beneficial to bring it to the fridge? Because the argument always needs to be what gives you a benefit? Yeah. Because I would keep

12:22Unidentified speaker (UNKNOWN): as much as possible

12:23Unidentified speaker (speaker_1): at room temperature, uh at room temperature and only bring down what it what it helps most. And

12:31Unidentified speaker (UNKNOWN): to

12:31Unidentified speaker (speaker_1): to bringing this let's say this shaping of the of the frequency, so

12:37Unidentified speaker (UNKNOWN): uh uh

12:38Unidentified speaker (speaker_1): to say it in this way, d bringing this continuous wave signal to room temper uh from room temperature to cryo and then do the shaping at cryo. That's a good for example, a good partitioning, yeah. As as uh as I said, Delft and and Intel showed it to bring also other showed it up to now but they showed it first. To do this at a four Kelvin level gives you a huge benefit because you save a more uh huge cabling effort to the four Kelvin stage and from there you can easily distribute it. That's not solving every issue, but that's a very nice step towards scaling. Because then you do the crucial or you do you do this frequency generation at room temperature, and distributing is also in this sense a little bit beneficial because you know better what you get at the cryotemperature. cryo temperature, yeah? Because that's one of the issues. The signal quality. and the control, how good the signal is, is much easier close to the qubit. That gives you also some performance issue uh performance advantages, for example.

13:51Unidentified speaker (speaker_0): Okay. And and the wave distribution, what exactly is that? Is it the DAC or the multiplexer, Plexor or is it the uh the set control or the couplers or the readout?

14:02Unidentified speaker (UNKNOWN): What

14:02Unidentified speaker (speaker_0): exactly?

14:03Unidentified speaker (UNKNOWN): Yeah, that

14:03Unidentified speaker (speaker_1): is mainly yeah, they uh what what you what you do is you are using uh the principle which is already used since decades in mobile phones, yeah. You just take one single frequency, so what you get from room temperature, and then mix it up with other frequencies. So you change the frequency on one hand side and the other hand side you can shape it, yeah. So you can make out of this continuous wave uh

14:30Unidentified speaker (UNKNOWN): uh

14:30Unidentified speaker (speaker_1): continuous wave signal, so which is a sinusoidal signal from which is not stopping. You you you get it the shaping, which means you start and stop it in a controlled way, and at the same time change the frequency a little bit by mixing it up or down uh from the center frequency, just to generate the right frequency for the one qubit. Then you can do, for example, that's frequency frequency multiplexing, for example. That's a different way of that's a different way of multiplexing. That's not distributing signals. Uh physically, it's distributing signals over the frequency domain. Yeah, so you generate several frequencies at the same time out of one frequency. So that's something that's a different way, a little bit different way of multiplexing. It's still called multiplexing, but in this case it's frequency multiplexing

15:21Unidentified speaker (UNKNOWN): then.

15:22Unidentified speaker (speaker_0): Okay, thank you for explaining. Let's maybe go into the product of iCirc. What is the role of RISC-V processor with SRAM and peripherals?

15:37Unidentified speaker (speaker_1): One you you mentioned in between, we did not go into the details about it, but latency also plays a certain role. Because you might know some qubit types, their coherence time, so the the the time during the these qubits are stable is very short, or let's say relatively short. And for this you need very low latency. So if you if you lose time just transferring signals to the top and bring it down, do there some processing. So you need to save some time, which means some of the controls, really digital controls, what need to be done, which waveforms you would like to have, whatever whatever you can think about. You can you can use, for example, RISC-V processors, so local processors at the cryo level.

16:26Unidentified speaker (UNKNOWN): And

16:26Unidentified speaker (speaker_1): RISC-V is the advantage. Because we would like to avoid royalties. No, but that's only it's finally an open source platform you can make use of. And finally the processing, you can especially you can you can uh specify it or or

16:50Unidentified speaker (UNKNOWN): uh

16:51Unidentified speaker (speaker_1): make it make it specific for your uh for your use case.

16:56Unidentified speaker (UNKNOWN): So

16:56Unidentified speaker (speaker_1): these RISC five are a very nice alternative as for a free usable, let's say, processor core. But a risk five is nothing else than a processor core where you can make use local control with, for example. Yeah. You can also make UCL Eucal processing, calculate some things and so on. So it's just a local local controller which helps

17:16Unidentified speaker (UNKNOWN): you

17:17Unidentified speaker (speaker_1): to I would say the main benefit is to make things faster and more efficient.

17:25Unidentified speaker (UNKNOWN): Okay.

17:25Unidentified speaker (speaker_0): Let's I went a little

17:27Unidentified speaker (UNKNOWN): bit

17:27Unidentified speaker (speaker_0): too fast to my next question. Let's go a step back. And uh how if we think about the heat

17:37Unidentified speaker (UNKNOWN): and

17:37Unidentified speaker (speaker_0): the allocation of the heat. how do you determine what stays at 300 K, what can move to 34k, and and what has to sit close to the qubits smaller than one millikelvin and

17:55Unidentified speaker (speaker_1): partitioning really is one of the crucial parts at the moment. As I stated before, I would not bring any anything into the cryostat which is not helping me there. And that's the same discussion you have on which temperature level you go. If you have a certain functionality which minimum which needs a certain power to get the performance and so on, you you cannot go down further than four Kelvin most likely, or to go to the three, four Kelvin regime, because the the heat load.

18:26Unidentified speaker (UNKNOWN): at

18:27Unidentified speaker (speaker_1): at the lowest Kelvin stages and that's again a little bit dependent on the technology of the qubit for example. But if we go for example for spin semiconductor qubits, we placed biasing circuits at the lowest Kelvin level and we demonstrated this working and it has had a certain benefit as I said, you make out of two cables a out of twelve cables you make two cables suddenly. Yeah. And these components you can afford at lowest temperature levels. And functions like this RISC V processor, what we mentioned before, where you step too fast forward, that's something I would never place at the milli Kelvin level because that's the the problems you generate by this is more than a benefit. So this is a typical component I would always place at a four Kelvin level. Also, how to generate all the power supplies, so the voltages you need in the system, I would also place this at a four Kelvin level and not at the lower levels. At the lower levels really the very

19:32crucial parts need to go. For example, when you make shuttling control in a spin qubit, for example. which is one possible technology which is very promising. Where you really need the quality and where you can avoid more heating by reducing the number of cables than you generate on your own. Then you need to go down. But that's that's a very, very complicated question. And that is really state of the art science to my opinion. How this will be solved in future, and that's very strongly dependent on technologies.

20:10Unidentified speaker (UNKNOWN): And

20:10Unidentified speaker (speaker_0): which which functions do you see most likely to move? Where do you see you said okay? It depends on the benefits that it brings. Where do you see the biggest benefits first?

20:24Unidentified speaker (speaker_1): Dependent on technology. I take now for example, the example I take the example of semiconductor based qubits. Biasing. They need a huge number of bias voltages. This need to move down, need to be moved down because you need to save the cables. For example, in to get nice noise free or noise reduced voltages, this need to be this need to be moved into the cryostat, that's clear. And uh even if we we are also working, that's more now on the institute level, but nevertheless, I bring in this experience in addition. There's this typical uh the for superconducting qubits, for example, you have this flux control, so where you finally tune the qubit a little bit in the frequency, where you can operate it and so

21:16Unidentified speaker (UNKNOWN): on. That

21:17Unidentified speaker (speaker_1): is also a huge benefit to bring it into the cryostat, but most likely on the four Kelvin stage, because superconducting qubits are so sensitive to temperature changes, but still you have a huge benefit because

21:32Unidentified speaker (UNKNOWN): in

21:32Unidentified speaker (speaker_1): this case it's mainly the quality of the signal which brings you the benefit. It's much easier, you can reach much higher qualities of the signals than you can in principle, or it's a huge effort to generate this benefits from room temperature, yeah, because then you need to have a pre-distortion. So you need to shape the signal in a certain way to compensate for things which are non ideal on the way down and so on. That makes makes

22:01Unidentified speaker (UNKNOWN): uh

22:01Unidentified speaker (speaker_1): operation much easier. for example. That would be the benefit for a superconducting qubit. And so you can discuss it for every technology in a different way, and even Ion trap companies are today looking into how they can bring DACs, for example, closer to their kubits, qubits because they also assume it will be a performance benefit. Yeah. And uh so partitioning and different what are the different trade-offs? That's a very, very complex discussion which is far from b far away from being already solved in this sense. And that we have to that we have to act now step by step

22:45Unidentified speaker (UNKNOWN): on.

22:46Unidentified speaker (speaker_0): Okay. And and what is the role that ISER ISERC can play in those areas?

22:53Unidentified speaker (speaker_1): To bring into these startups and scale up the knowledge what is reachable by semiconductor components. And then really to make the right decision what could be a next step and then develop this component for the companies and with the companies. and then coming closer to their target, most likely scaling, for example. But can be also pure performance improvement, for example, is also possible. Which they need to further develop.

23:36Unidentified speaker (UNKNOWN): Okay,

23:37Unidentified speaker (speaker_0): fair enough. Let's go. Let's see. I think we asked I asked most of the questions regarding the

23:51Unidentified speaker (UNKNOWN): which

23:51Unidentified speaker (speaker_0): which functions to move

23:53Unidentified speaker (UNKNOWN): to

23:53Unidentified speaker (speaker_0): lower temperature stages, or did I miss uh a a function or a building block that has is has to go to lower. to stages.

24:06Unidentified speaker (speaker_1): Yeah, one one crucial point is uh let's say what is what is still a little bit open for me, but I think there will be solutions. This fault tolerant topic, for example, where you where you need to extract which which things went wrong and where do you need We to correction and so on. But I think there are there are fast connects or broadband connections available. You need to do some let's say there are in principle technologies available anyhow, for example an optical site. But them let's say we are also a little bit working on electrical solutions for this to make it very efficient and very fast possible. This is something which is a little bit open to my opinion, but to to really, for example, bring the fault tolerant main function, so the decoder to cryo

25:10I d that for me that that looks not that it can be really beneficial. because the processing power you need for this is most likely too high to bring it to the cryostat. But that's some people some people claim okay we can do it at cryo, okay. I'm I'm open to see what is happening there, but uh whatever way you go, there are there are still huge issues you have to you have to solve in in the in one way or the other. And definitely also from a semiconductor point of view, a company like ICER can can work on all these things, but that need to be clarified first on system level what is most likely the beneficial side, which is let's say one or two hierarchy levels above what we what we normally tackle. That's really on the highest higher higher system level.

26:01Unidentified speaker (speaker_0): And and and what you mentioned, Stefan, is is is the is that a rapid feedback look for quantum error corrected qubits?

26:08Unidentified speaker (UNKNOWN): Yes.

26:09Unidentified speaker (speaker_1): Or

26:09Unidentified speaker (UNKNOWN): is that

26:09Unidentified speaker (speaker_0): something that's

26:11Unidentified speaker (UNKNOWN): no

26:11Unidentified speaker (speaker_1): no, that is what I meant. That is mainly what I

26:13Unidentified speaker (UNKNOWN): meant.

26:14Unidentified speaker (speaker_0): Okay. Yeah. Okay. Can we talk can we zoom in into the architecture of uh iCirk? Circ. We we touched a little bit a little bit based on it with the RISC-V RISC V processor.

26:26Unidentified speaker (UNKNOWN): Um

26:27Unidentified speaker (speaker_0): that effectively becomes the the cryogenic controller. And you explained okay why you have chosen risk V because that's that's an open model and you don't have to pay the high royalties that you have to pay to ARM. Aside from that, does RISC V given its do they have more functions or is the functionality the same? And potentially more. function because it's open uh it's open it's open model.

26:56Unidentified speaker (speaker_1): In principle it's the same functions, but you can customize it easier. So if you need customization, it's in principle possible. But that's let's say that's not that's not the stage I would say the field is to think about this customization in that sense. And if we think about customization we may mainly think about how we can make it more efficient or how we can reduce a little bit power consumption. But due to the fact that we place it at higher stages for Kelvin, for example, that's not the crucial part at the moment. But to have a local control sometimes can help you to be faster, as I said before. And that is for example also what we are just evaluating in our in s in my institute to see what is could be the benefits there. But in principle the technology is ready to bring into the cryostat. And also that is depending on on customers and and little bit architectures on on customer side, if it could be beneficial or not. Yeah. But it's a t as technology it is available. It is in principle demonstrated that it is working and so it's uh

28:03Unidentified speaker (UNKNOWN): that

28:04Unidentified speaker (speaker_1): just a just an additional tool you have in your toolbox now to to play around with a little

28:10Unidentified speaker (UNKNOWN): bit. And

28:12Unidentified speaker (speaker_0): you mentioned okay, you it's it's easier to standardize in RISC V. What would you try to customize and what would you try to scrap versus a the conventional ARM ARM-based based processor to optimize specifically for quantum control?

28:28Unidentified speaker (speaker_1): Yeah, especially if if we are thinking again about this error correction topic, there might be additional some some operations locally needed perhaps which which are difficult to

28:41Unidentified speaker (UNKNOWN): uh

28:42Unidentified speaker (speaker_1): do in the standard uh in the standard set of opera uh uh or in the the standard set

28:48Unidentified speaker (UNKNOWN): uh

28:48Unidentified speaker (speaker_1): what you can do with a processor but that's uh let's say that's a little bit looking into the future and looking in into a into a wall out of fog where you do not really see what is how it how it's going in in future. But this customization could be useful to have special commands for example, yeah where you can easily do not need uh spend too much uh the memory on or where you can transfer a certain set of data. easier to it up or perhaps even to do some decoding. But perhaps even what just might be more beneficial is to do some corrections. Yeah, because the corrections need to be done locally in a certain way. Because error correction always means not to collect what is the error, also try to correct things on the lower level and that is in a local loop most likely easier and there you might need special commands which would take too long if you are not if you if you need to do it in a more sequential way or whatever, yeah. But

29:48that's that's really on a let's say that's on a high level of speculation what what might be the benefit there. Yeah. That has to turn out definitely. But it gives you some flexibility and that helps you already.

30:02Unidentified speaker (UNKNOWN): Okay.

30:03Unidentified speaker (speaker_0): And the second component is so it was RISC-V,

30:06Unidentified speaker (UNKNOWN): and

30:06Unidentified speaker (speaker_0): it was SRAM. Why S RAM? And in

30:13Unidentified speaker (UNKNOWN): SRAM

30:13Unidentified speaker (speaker_0): SRAM

30:14Unidentified speaker (UNKNOWN): is

30:15Unidentified speaker (speaker_0): in AI it's very fast memory and it's small. But how does S RAM uh compare on energy usage and and on heat?

30:26Unidentified speaker (speaker_1): Yeah that's a that's a good question. I learned not too long ago, I think it was two weeks ago that that there are some some people also working on on special DRAM technologies now in the cryostat. RAM expert, But for me the question is it could be the case that DRAM is perhaps also a bit beneficial, yeah? Because if you need if you need bigger because SRAM you are always limited how much how much uh memory do you have available in a certain way, yeah. And a DRAM could be beneficial and DRAM could benefit also from the experience what we have in cryotemperature, at cryotemperatures, that for example leakage is not as high as as we know it from room temperature. And that's why one of the let's say perhaps DRAM can can benefit from it, but I have some experience which is quite quite some years ago when we had first discussions with memory companies. Some people said DRAM will never work at Cryo. Okay, but that to be honest, that's ten years ago.

31:35But that could be one of the crucial points, especially in we we are just preparing one publication together with our partners, from Institute, but also from iCerc, but also from Arc Systems, for example. And there this memory question is one of the crucial points. Because we need local memory and not let's say we we need to finally distribute it, let's say it in this way. We need distributed memory which is local and not in one memory block, for example. And then the discussion: can you do this with DRAM, SRAM? How can you do this? That is that's one crucial point also to get this power consumption issue what you mentioned, but also uh to to get to get it done in a certain certain way. how can you optimize the cost on this point of view? Because what we are looking in there is for example clear scaling beyond ten thousand. Yeah. That's and and then you run into cost at a certain point you run into cost issues. And

32:41even there it could be in a discussion if an local memory gives you a problem on cost or if it is even beneficial, that's unsolved at the moment. That's not really. not thought through from the beginning to the final end at the moment.

32:56Unidentified speaker (speaker_0): Stephen, you mentioned you need a local memory and a distributed memory. Why is

33:01Unidentified speaker (UNKNOWN): that?

33:03Unidentified speaker (speaker_1): Because transfer let's say that's one on one hand we we have a latency, still this latency issue, yeah. Because to to read out the memory out of a central memory, always we w need energy. and time. And both things at some point you cannot afford, yeah. And to have it locally available gives you a certain but that's the same discussion as you have it in AI, yeah. To be honest, AI is also mainly uh or is a huge disadvantage that you always have this this memory bottleneck, yeah. And this that's finally the same. Because the quantum computer finally, due to this coherence time issue, so the the limited lifetime I would call it of the of the qubits, you need to be fast. Time is one of the most crucial points. And on the other hand, on the other hand, power is the second crucial point. This easily counts for local memory. yeah. But it's a question if you can afford it from a from a cost perspective. point of view. And yeah that's

34:06Unidentified speaker (UNKNOWN): But

34:06Unidentified speaker (speaker_0): what what Stefan,

34:07Unidentified speaker (UNKNOWN): what do

34:07Unidentified speaker (speaker_0): you mean with local memory? Because if it's in the fridge it's very close already. The distance is not like a few meters.

34:15Unidentified speaker (speaker_1): Yeah that's that's that's true, but we would like to have it even more local. Local to Z

34:21Unidentified speaker (UNKNOWN): the

34:22Unidentified speaker (speaker_1): two that's say we we have a very we we have a very modular architecture what we are developing. And modular means you you just would like to in a in scaling it up from a chip point of view, you would just like to use it as a unit cell, yeah, and copy it step by step. And if you would have then a memory which is not in this in this unit cell, you would need to have an addition to th the then you cannot really make it as a let's say as a copy and paste scaling. You would need then always have access to an external memory, for example. And that I I really talk about local local memory and not only local memory. I'm even talking about local square memory, yes.

35:08Unidentified speaker (UNKNOWN): Which is a challenge.

35:11Unidentified speaker (speaker_1): Yeah, we really in every

35:13Unidentified speaker (UNKNOWN): office. of this in

35:15Unidentified speaker (speaker_1): every of this unit cells have their own memory cell which is then once programmed from outside and or set in the set or even generates the data internal in the local cell and uh is then stored there so that the access of this unit cell is directly to the to its own memory.

35:37Unidentified speaker (UNKNOWN): But

35:37Unidentified speaker (speaker_0): is this is this also in full involving in memory compute?

35:42Unidentified speaker (UNKNOWN): Or

35:42Unidentified speaker (speaker_0): not?

35:42Unidentified speaker (UNKNOWN): Yeah

35:43Unidentified speaker (speaker_1): that's even that's even that's one of my things also my my institute is doing in addition. We call it neuromorphic computing, which is close to this memory co compute in memory, for example. Yes, could be on the long run even including this, yes. Could be, but uh I'm very that's very far away at the

36:04Unidentified speaker (UNKNOWN): moment.

36:06Unidentified speaker (speaker_0): What is interesting in AI you have C CPU, GPU, but in the future the the for me the the holy grail im

36:15Unidentified speaker (UNKNOWN): I

36:16Unidentified speaker (speaker_0): I went to a memory conference in in in a workshop in Leuven and the at that point of time all the the memory professors they said the the holy grail is in memory compute.

36:29Unidentified speaker (UNKNOWN): I

36:29Unidentified speaker (speaker_0): was thinking well if you talk about in memory compute

36:33Unidentified speaker (UNKNOWN): you

36:33Unidentified speaker (speaker_0): talk about quantum computing. So

36:40Unidentified speaker (speaker_1): okay, that's an inter that's an interesting view on this, yes.

36:44Unidentified speaker (speaker_0): So so in a way if you have a CPU, a QPU and a Q uh uh a GPU and a QPU can the Q QPU maybe contribute to the in memory computing item? That's maybe too far, but

36:56Unidentified speaker (speaker_1): yeah, but that's an interesting you you see all the fields are th they are not independent of each other. And as I said, my institute is working also in this field more on a let's say on a circuit and and and l let's say component level. But we also already thought about if this compute in memory could help us there. But let's say that's too uh I would say that we can think about this in some years about if it if it gives us further benefit. But and then also the field ev evolves a little bit more about this. But in principle,

37:32Unidentified speaker (UNKNOWN): when

37:32Unidentified speaker (speaker_1): I say local local memory, that's not far away from this. Yeah.

37:37Unidentified speaker (speaker_0): Okay.

37:37Unidentified speaker (UNKNOWN): How

37:38Unidentified speaker (speaker_0): much SRAM do you ultimately need per thousand

37:41Unidentified speaker (UNKNOWN): to per

37:42Unidentified speaker (speaker_0): million qubits?

37:43Unidentified speaker (UNKNOWN): And

37:44Unidentified speaker (speaker_0): does SRAM area or power eventually become a scaling bottleneck?

37:50Unidentified speaker (speaker_1): Yeah that's that's that's really one of this crucial questions which is open to my to my point of view. And that let's say that's first strongly depending dependent on the qubit technology in the fields we are mainly active in is it can be really but that is the reason that we say okay we would like to have it more on the local side that it is not becoming a bottleneck in that sense for scaling. But it could be in principle but even that is perhaps from a cost point of view it's not clear if it is possible. But let's say the cost issue, not so many people are talking about this at the moment, but it could be. But at the moment we do not think that's a scaling bottleneck, because all other functionalities we need, for example, for such a unit cell in a silicon based pin qubit quant computer,

38:50Unidentified speaker (UNKNOWN): the

38:50Unidentified speaker (speaker_1): S RAM will be most likely not the bottleneck for scaling because that's not our limiting factor in if we think about S RAM. It uh just assume that is not our limiting factor in form factors or in size at the moment. And due to this is most likely not the limiting factor in in from a from a cost point of view at the

39:12Unidentified speaker (UNKNOWN): moment.

39:13Unidentified speaker (speaker_0): Stefan, what is the we're ramping up the interview? As the biggest constraint

39:24Unidentified speaker (UNKNOWN): for

39:24Unidentified speaker (speaker_0): Cryosimos.

39:27Unidentified speaker (speaker_1): From external is acceptance that people really need it. That's one constraint. So we need to bring it into the systems and the market and there we need some examples for. And on the long run, I hope this will be not an issue is availability of this technology. Because this tep technology, okay, we have it in Europe on S T side, but as most of the companies are working with global foundries, our constraint hopefully will never become that we not do not have availability. Even it is produced in Dresden at the moment. But you never know, yeah? On the on the global side you never

40:03Unidentified speaker (UNKNOWN): know what happens.

40:06Unidentified speaker (speaker_1): Yeah, exactly. We are talking about FDSOI. But that is what we are talking also, which is available at at ST, but that is not the technology we worked with. And that's a little bit different size, but it's not let's say that's not the crucial difference. Let's say at a certain point it will be a limiting factor, how far can we scale with this type of cryosemus? So do we need at a certain point to go down in feature size, for example? Or do we need even a different technology? That is what I meant with Semicon has this real cool technology available. Do we need to step in such a direction?

40:50Unidentified speaker (UNKNOWN): But

40:50Unidentified speaker (speaker_1): it's from a cryosemos point, we are far away from this these limitations at the moment. So that's that's we are from a let look into this. This cryosemos business is still so new that we like I would say uh I'm not sure if we are not too early with founding the company because two years ago there was no not there was no market in this field. Absolutely nobody was interested in it, except of some crazy guys like us who worked on this. And this is changing now, but we are in a such early stage that's uh that's we are far far be before these limitations. But we need to think sink in into it because uh sometimes developments are faster than time, yeah. So what you need to prepare things. And so that's the reason for us and that's also important for us as a company. We have still in the backbone this scientific activities which are real scientific, which have has no interaction on the on the

41:59operational side with the company. But this scientific view looking into what are future directions, what is going, that's finally also beneficial for us, yeah, because this there will be further development and there will be the further evolutions and even perhaps revolutions on this side. And uh but let's do it as a company step by step. And then we establish establishing in the market and establishing the market first. That's our first goal now. And then we are also definitely we are not we are not ignoring what is happening in the future, but uh nevertheless we need to act step by step now.

42:42Unidentified speaker (speaker_0): Stefan, my last question if we meet again in in twenty thirty, what single uh technical milestone would make you say

42:51Unidentified speaker (UNKNOWN): uh

42:51Unidentified speaker (speaker_0): the control electronic problem uh essentially

42:55Unidentified speaker (UNKNOWN): uh

42:55Unidentified speaker (speaker_0): that is solved?

42:58Unidentified speaker (speaker_1): Yeah w let's say twenty thirty so that's around about three years. No it's four years. That's not important. So if we have a scaled computer, we see a scaled computer in a let's say in the solid state domain, so semiconductors or superconducting, with scaled beyond this I would say this 10,000 qubit limit then I would say then we reached a really major milestone and that will definitely not happen without cryosemos.

This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3+sortformer-v2.1 and interpreted by Qwen3.6-35B-A3B-UD-IQ3_S.gguf. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.
All insights