Skip to content

Meeting 2026 07 13

amirshehataornl edited this page Jul 18, 2026 · 5 revisions

Meeting - July 13, 2026

Upcoming meetings

Agenda

  • Speaker: Anastasiia Butko (LBL) overview on recent work related to control electronics at LBL

Attendance

Person Institution
Amir Shehata ORNL
Michael Ferguson HPE
Bran Radovanovic ORNL
Thomas Naughton ORNL
Andy Stone HPE
Alain Roy IonQ
Eun Kyung Lee IBM
Claudio Siqueira de Carvalho IBM
Jeff Heckey AWS
Smriti Bajaj Dell Technologies
Brandon Neth HPE
Nick Wright NVIDIA
Namit Anand HPE
Miwako Tsuji RIKEN
Alex Chernoguzov Quantinuum
Yasuko Eckert AMD
Jonathan Skone NERSC
Ryan Landfield ORNL
Neal Erickson Quantinuum
Christian Ortiz Quantum Brilliance
Kevin Kissell Alice & Bob
Anastasiia Butko LBNL
Josh Moles IonQ
Abdur Rahman Hatim IISc
Martin Schulz TUM

Notes

  • NOTICE: Notes generated using AI; please verify accuracy.

1. Opening

  • Thomas Naughton started the recording and confirmed that Anastasiia Butko would be presenting on control electronics work at Berkeley Lab.
  • Anastasiia introduced herself as part of the Berkeley Lab computer architecture group, with a background in classical architectures, microelectronics, HPC, and computer science.
  • She framed the talk around the quantum hardware stack, especially control challenges for superconducting qubits.

2. Presentation Overview (Anastasiia Butko)

2.1 Quantum Computing as a Full Hardware Stack

  • Anastasiia emphasized that a quantum computer is more than a quantum processor:
    • The quantum chip is only one part of a larger system.
    • The classical control stack, cabling, dilution refrigerator, readout, and feedback path are all part of the machine.
  • The Berkeley Lab quantum testbed was used as an example of this full-system view:
    • A small superconducting chip with roughly a dozen qubits still required substantial room-temperature control hardware and wiring into the refrigerator.
    • Commercial control hardware was available from vendors, but Berkeley Lab also developed in-house FPGA-based control stacks.
  • She described in-house work around open-source control components, including Quasar and Cubic.
  • A major practical motivation was access and flexibility:
    • Vendor hardware could provide comparable quality.
    • In-house control gave researchers detailed access to each layer of the control hardware and software.

2.2 Superconducting-Qubit Control Constraints

  • Superconducting qubits were presented as one of the more mature quantum technologies, but with strict control constraints.
  • The key example was mid-circuit measurement:
    • A circuit may need to measure a qubit partway through execution.
    • The control system then branches conditionally based on the measured state.
  • These workflows are latency-sensitive because:
    • Superconducting-qubit coherence times are limited.
    • Single-qubit gate pulses may be on the order of 10 to 20 nanoseconds.
    • Two-qubit gates and readout operations are longer.
    • Feedback, signal travel, and control processing time all consume part of the available circuit window.
  • Anastasiia contrasted two control approaches:
    • A simpler offline model where software compiles an algorithm down to a fixed pulse sequence and the room-temperature controller plays it back.
    • A more dynamic model where a quantum control processor sits closer to the hardware and can generate pulses and make decisions during execution.
  • The dynamic model can reduce latency and support more flexible circuits, but it requires specialized hardware and is harder to implement and modify.

2.3 Quantum ISA and RISC-V-Based Control

  • Anastasiia described earlier work on a quantum instruction set architecture for control processors.
  • The ISA work examined classical architecture questions such as:
    • Instruction encoding width.
    • Whether instructions should be vectorized.
    • How many bits are needed to address qubits.
    • How those choices limit simultaneous control and scalability.
  • The design used a lightweight RISC-style approach rather than a complex instruction set.
  • A RISC-V-based implementation was attractive because:
    • The ecosystem already includes compilers, software stacks, and hardware implementations.
    • Existing tooling can be extended rather than rebuilt from scratch.
  • Berkeley Lab implemented a processor on a VC707 FPGA connected to the testbed and demonstrated fast feedback on real qubits.
  • The same control stack was also used for measurements such as T1 characterization, showing that the work was not only simulation-based.

2.4 State Discrimination and FPGA AI Engines

  • The second major topic was the quantum state discriminator, which classifies readout data into qubit states.
  • Anastasiia described the readout input as IQ-plane data or blob-like readout distributions:
    • A simple threshold line can work when the readout blobs are well separated.
    • Overlapping or closely spaced blobs require more sophisticated discrimination.
    • Neural-network-based discriminators have been explored for improved accuracy and performance.
  • Cubic ML, developed by Berkeley Lab colleagues, was used as a baseline:
    • It implemented a neural network on FPGA programmable logic.
    • It could support fast state discrimination, but scaling is limited by FPGA resource use and synchronization issues.
  • Anastasiia explored using the Xilinx VCK190 AI-engine architecture:
    • The AI engine provides an array of specialized vector-processing tiles next to the programmable logic.
    • She mapped a three-layer neural network to the AI engine.
    • The AI-engine version had about 81 nanoseconds of kernel inference time compared with about 54 nanoseconds for Cubic ML.
    • Although slower for the small network, it stayed under 100 nanoseconds and may scale better for more qubits or larger neural networks.
  • Resource and power measurements were included because power becomes a major constraint when control electronics move into low-temperature environments.

2.5 Low-Temperature and Superconducting Control

  • The final part of the talk asked whether parts of the classical control hardware can be moved from room temperature into the refrigerator.
  • The motivation is to:
    • Reduce feedback latency.
    • Reduce wiring complexity.
    • Improve scalability beyond near-term experiments.
  • Anastasiia identified several challenges:
    • Electronics near qubits can introduce noise.
    • Standard FPGAs and CMOS chips cannot simply be placed in millikelvin environments.
    • Cryo-CMOS is easier to adopt because it remains CMOS-like, but may have unacceptable noise and power characteristics.
    • Non-CMOS superconducting technologies require architectural redesign, not only technology substitution.
  • She discussed superconducting classical logic technologies such as RSFQ and AQFP:
    • These technologies can operate at very high clock rates.
    • Existing computing work in this area predates current quantum-control applications.
    • Mapping conventional designs directly to these technologies can be impractical.
  • A neural-network state discriminator was used as an example:
    • Directly mapping the network to superconducting logic required about 344,000 Josephson junctions, which Anastasiia described as impractical.
    • Redesigning the representation and operations with a biologically inspired unary data representation reduced the design to roughly 3,000 Josephson junctions.
    • The redesigned implementation was described as very fast and low power.
  • Anastasiia said a Nature paper on this low-temperature discriminator work had been accepted and was moving through publication processing.

3. Discussion & Q&A

3.1 Low-Temperature Data Movement and Pulse Generation

  • Amir Shehata asked whether low-temperature control still needs to send data back to room-temperature compute and receive commands from outside the refrigerator.
  • Anastasiia clarified that, for the low-temperature state discriminator work, the readout classification does not need to leave the low-temperature environment.
  • She noted that other pieces of the control stack are still needed, especially memory-related pieces, which remain challenging for superconducting implementations.
  • Amir asked whether only state discrimination had moved to low temperature, not pulse generation.
  • Anastasiia said the team also had experimental work with a simplified Quasar-like core through the same tool flow, suggesting that additional low-temperature control pieces may be feasible.
  • She cautioned that pulse-table handling and on-the-fly pulse generation remain research problems rather than solved engineering components.

3.2 Relationship to Other Open-Source Control Work

  • Jeff Heckey asked about the relationship between Berkeley Lab's control work and a Fermilab control-board project.
  • Anastasiia distinguished the Fermilab work from Berkeley Lab's Cubic work:
    • They are separate projects.
    • Both are open-source efforts.
    • They developed in parallel for somewhat different control contexts.
  • She noted that there had been interest in adding a RISC-V processor into that broader ecosystem, though she was not sure of the current status.

3.3 System Programming and Integration

  • Jeff also asked about programming the AI-engine tiles and managing fan-out across multiple control systems.
  • Anastasiia said this remains difficult to unify and is still largely handled by working at separate levels and stitching components together.
  • A practical blocker is board capability:
    • The VCK190 board has useful AI-engine capabilities.
    • The GCU board used in the current setup has a different DAC/ADC configuration.
    • Using two boards would create significant board-to-board data-movement complexity.
  • Anastasiia said they were waiting for a future board that combines the relevant capabilities and would enable cleaner experiments.

3.4 Paper and Slides

  • Josh Moles asked whether there was a preprint or publication timeline for the accepted paper Anastasiia mentioned.
  • Anastasiia said there might be an arXiv version and invited Josh to email her so she could follow up with the paper.
  • Nick Wright dropped a link to a related AI-engine paper during the discussion.
  • Thomas asked Anastasiia to share her slides with him or Amir, and she agreed.

4. Key Takeaways

  • Quantum-control hardware is a central part of the quantum-computing system, not a peripheral implementation detail.
  • Superconducting-qubit control is strongly constrained by latency, especially for mid-circuit measurement and conditional branching.
  • RISC-V-based control processors and FPGA-based state-discrimination hardware offer practical near-term paths for more flexible control.
  • AI-engine accelerators may provide better scaling for neural-network-based state discrimination even when small single-qubit examples do not show the lowest latency.
  • Moving control closer to the qubits is attractive, but low-temperature and superconducting implementations require new architectures, memory strategies, and careful noise and power management.
  • Integration tooling and board-level system design remain major practical barriers for experimental control-stack work.

5. Action Items

Action Owner Notes
Follow up with Anastasiia about the accepted low-temperature state-discriminator paper or preprint. Josh Moles Josh offered to email her; Anastasiia said she could share or point him to the paper.
Share slides from the presentation. Anastasiia Butko Thomas asked Anastasiia to send slides to him or Amir, and she agreed.

6. Closing

  • Thomas thanked Anastasiia for the talk and apologized for the earlier room or meeting-access issue.
  • Anastasiia said she was happy to share the materials and thanked the group for having her.
  • Thomas closed by thanking everyone and wishing attendees a good rest of the week.

Files

Clone this wiki locally