Skip to content

Meeting 2026 08 17

amirshehataornl edited this page Aug 20, 2026 · 4 revisions

Meeting - August 17, 2026

Upcoming meetings

Agenda

  • Seminar Series: Infleqtion - Ryan Jones

Attendance

Attendance Report was corrupted. The meeting had 42 attendees.

Person Institution
Amir Shehata ORNL
Ryan Jones Infleqtion
Johannes Blaschke EIT Oxford
Masoud Mohseni HPE
Namit Anand HPE
Jiri Schindler IonQ
Peter Groszkowski ORNL
Álvaro Caride IQM

Executive Summary

  • Ryan Jones presented Infleqtion's neutral-atom quantum-computing stack and discussed how its hardware capabilities affect quantum-HPC integration.
  • The talk emphasized dynamic atom reconfiguration, all-to-all connectivity, and how neutral-atom systems can support different fault-tolerant and measurement-feedback strategies.
  • Infleqtion's Superstaq middleware was described as the main user-facing interface for compiling circuits from common quantum programming frameworks to native hardware operations.
  • Ryan distinguished batch-time hybrid workflows, where quantum and classical steps may run minutes or longer apart, from real-time workflows where classical compute participates in the loop with the QPU on roughly millisecond timescales.
  • The discussion focused on when co-location or co-allocation with HPC resources matters, how real-time decoder and conditional workflows could be exposed to users, and whether MLIR-style heterogeneous kernels could provide a standard interface across vendors.
  • Participants raised questions about mid-circuit measurement, conditional operations, QLDPC decoding requirements, repeat-until-success circuits, RoCE/RDMA-style links, NVLink-related integration, and FPGA-based real-time data movement.

Notes

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

1. Opening

  • Amir Shehata opened the meeting and introduced Ryan Jones from Infleqtion for a seminar on QHPC integration.
  • Ryan introduced himself as a principal embedded systems engineer at Infleqtion, focused on control software and hardware architectures for quantum computing.
  • He invited questions during the talk through raised hands or chat.

2. Presentation Overview (Ryan Jones)

2.1 Infleqtion and Neutral-Atom Systems

  • Ryan described Infleqtion as a quantum-technology company centered on neutral-atom systems.
  • The same core neutral-atom platform supports:
    • Research, academic, and industrial neutral-atom cells.
    • Position, navigation, and timing systems.
    • Sensing applications such as inertial sensing.
    • Quantum-computing systems and quantum software.
  • Infleqtion has offices in the United States, United Kingdom, and Australia, and systems deployed with collaborators internationally.
  • Examples included quantum computers delivered to the UK's NQCC and Japan's Moonshot project, PNT systems, and space-related atom-lab work.

2.2 QPU Capabilities

  • Ryan framed neutral atoms as attractive qubits because they are naturally uniform, stable, scalable, and have relatively long coherence times.
  • He noted that longer coherence times can relax some latency requirements for online decoders compared with faster-clocked modalities.
  • Infleqtion's system uses laser beams, magnetic fields, electric fields, microwave radiation, and atom motion to manipulate qubits.
  • A central hardware capability is dynamic reconfiguration:
    • Atoms can be physically moved between sites.
    • Arrays can be prepared before circuit execution.
    • Reconfiguration can also occur during circuit execution.
  • This dynamic motion enables effective any-to-any or all-to-all coupling, which affects how users and compilers should express connectivity and motion costs.

2.3 Fault-Tolerant Operations and Measurement Feedback

  • Ryan used a fault-tolerant state-preparation example to show how atoms can be moved after two-qubit operations to continue forwarding quantum information.
  • For many quantum computers, error-correction workflows involve measuring qubits, sending the result to a classical decoder, applying corrections, and continuing execution.
  • Infleqtion's all-to-all connectivity may enable alternatives such as measurement-free quantum error correction and coherent feedback.
  • Ryan emphasized that users should still be able to use standard circuit tools, but should also be able to expose or hint hardware capabilities such as atom motion and all-to-all connectivity to the compiler.
  • He noted that, for very large circuits, it may sometimes be better to measure and use a classical decoder despite the latency cost, depending on the full system tradeoff.

2.4 Superstaq and Compilation Stack

  • Ryan described Infleqtion's high-level stack as:
    • An application layer where people and software work with collaborators to formulate problems.
    • Superstaq middleware as the interface between user-level quantum circuits and hardware.
    • A native gate and pulse-sequence layer that executes on the control system.
  • Superstaq was described as accepting common circuit inputs from frameworks such as Qiskit and Cirq.
  • The compiler can target Infleqtion neutral-atom systems and other third-party hardware, including IBM systems, trapped-ion systems, superconducting systems, and spin-qubit systems.
  • In response to a chat question, Ryan said Superstaq supports OpenQASM with some limitations, but is not fully feature complete.

2.5 Batch-Time Hybrid Workflows

  • Ryan defined batch-time workflows as cases where one circuit or a collection of circuits is submitted and eventually returns measurement data for later classical processing.
  • These loops may take tens of minutes, hours, or longer.
  • He used precision oncology work with the University of Chicago and MIT under Wellcome Leap's Q4Bio program as an example.
  • The workflow was described as:
    • Preprocess patient sample data.
    • Frame the problem as a polynomial constrained binary optimization tournament.
    • Use the quantum computer to identify important features and relationships.
    • Use classical computation to narrow the problem until the remaining instance is classically tractable.
  • Ryan argued that physical co-location matters less for these long-loop workflows if the classical and quantum stages already take many minutes.

2.6 Real-Time Hybrid Workflows

  • Ryan contrasted batch-time workflows with real-time workflows, where classical compute is inside the execution loop of the quantum computer.
  • A real-time workflow might:
    • Execute part of a circuit.
    • Measure the QPU state.
    • Send measurement data to a classical system.
    • Receive low-level operations back from the classical system.
    • Continue execution based on the result.
  • He suggested the relevant loop period is likely on the order of milliseconds, including both classical and quantum durations.
  • Ryan said these real-time workflows are currently internal to Infleqtion, but are worth discussing because standardization could eventually let users define classical operations inside the QPU loop.

2.7 Real-Time Hardware and Heterogeneous Kernels

  • Ryan described Infleqtion's real-time control hardware as distributed programmable systems-on-chip used to orchestrate peripherals, modulate laser beams, acquire measurements, and manage QPU state.
  • These systems are connected through a high-speed topology, with some links in the hundreds-of-gigabits-per-second range.
  • He described a RoCE/RDMA-style link for expanding real-time compute capabilities to accelerators such as GPUs.
  • The same dispatch model could in principle extend beyond GPUs to FPGA, TPU, MPU, DPU, CPU, or other accelerator clusters.
  • Ryan then described a heterogeneous-kernel model:
    • Quantum instructions execute on the real-time control system.
    • A measurement operation produces data.
    • The control system dispatches a classical decode function to an accelerator such as a GPU.
    • The accelerator returns a result with low latency and low jitter.
  • He suggested MLIR as a promising path for expressing these heterogeneous workflows in a way that different vendors could lower to their own hardware.
  • He mentioned open-source efforts such as the Munich Quantum Toolkit and PennyLane as examples of MLIR-related quantum tooling.

3. Discussion and Questions

3.1 Mid-Circuit Measurement and Conditional Operation

  • Masoud Mohseni asked whether Infleqtion's pipeline provides access to mid-circuit measurement and conditional operation for heterogeneous compilation.
  • Ryan distinguished between:
    • Mid-circuit measurement, which can be described in a circuit-style interface.
    • The heterogeneous real-time part, which is currently implicit and not publicly exposed to end users.
  • Ryan said Infleqtion does not yet have a sustainable public interface for users to describe arbitrary heterogeneous workflows.
  • He noted that such workflows could be described in OpenQASM-like or circuit-like forms, but that the full hardware degrees of freedom make a simple circuit-only abstraction less tractable.

3.2 Decoder Latency, Throughput, and QLDPC Codes

  • Masoud commented that neutral atoms' slower physical clock can relax decoder latency requirements, but may create throughput challenges for fully parallel transverse operations and conditional operations.
  • Ryan agreed that the earlier latency framing was simplified and that neutral-atom systems have challenges as well as advantages.
  • He said he was not the best person to describe Infleqtion's detailed approach to that particular throughput issue.
  • Namit Anand asked what class of QLDPC codes Infleqtion expects to support and noted that more complex QLDPC codes create decoding requirements that interact with Masoud's point.
  • Ryan said Infleqtion has parallel efforts across classes of QLDPC codes but could not describe a specific direction or family in detail.
  • He offered to discuss hardware-side approaches for reducing latency, but not the internal code-family roadmap.

3.3 Co-Location, Co-Allocation, and Scheduling

  • Amir asked about the difference between physical co-location and co-allocation of quantum and HPC resources.
  • He noted that even if a batch-time workflow tolerates high network latency, releasing and reacquiring HPC resources around QPU use could be expensive because of queue delays.
  • Ryan agreed that co-location may not be compelling from a data-transfer-latency perspective for long-loop workflows, but that co-allocation and orchestration can matter substantially.
  • He said a tightly coupled quantum/classical scheduling model is a reason an openQSE-style interface could be preferable to cloud-service workflows for some use cases.
  • Amir also asked whether Infleqtion uses workflow management software for batch-time workflows.
  • Ryan said Superstaq has internal workflow-management components, but they are not generally exposed to users.
  • He said Infleqtion is considering compatibility with QRMI-like interfaces so HPC orchestration systems can interact with Infleqtion systems and users can keep using familiar workflow managers.

3.4 Algorithmic Need for Real-Time Workflows

  • Masoud asked for examples of quantum algorithms that require real-time millisecond-scale workflows.
  • Ryan identified quantum error correction and decoding as obvious examples.
  • He also suggested repeat-until-success or early-exit conditions, where an algorithm can continue until a success criterion is reached or stop when a failure condition is detected.
  • Masoud clarified that he was asking about algorithmic quantum/classical co-processing at the logical layer, beyond error correction or error mitigation.
  • Ryan said some tournament-style workflows could benefit from faster feedback hooks, but did not provide a concrete public algorithmic example.
  • Amir said it would be useful for Infleqtion to share specific algorithms that require in-the-loop classical HPC beyond decoding.
  • Ryan agreed and said Infleqtion expects to use its low-level flexibility in creative application-layer workflows and to make some of those examples publicly available.

3.5 Repeat-Until-Success and Resource Estimates

  • Namit commented that repeat-until-success circuits appear not only in error mitigation but also in partially fault-tolerant computation and Clifford+T synthesis resource estimates.
  • He cautioned that reporting average T counts under a repeat-until-success assumption hides variance and implementation details.
  • Namit suggested either:
    • Expose repeat-until-success as an explicit primitive so users can declare it when submitting jobs.
    • Derandomize the construction and report an explicit circuit and approximation error instead of only an average estimate.
  • Ryan asked whether Namit meant these should be formal primitives; Namit confirmed that was the suggestion.

3.6 RoCE/RDMA, NVLink, and FPGA Data Movement

  • Jiri Schindler asked about the real-time hardware slide with the RoCE/RDMA-style connection, including what abstractions or libraries Infleqtion uses and whether the accelerator connection is actually RoCE or NVLink.
  • Ryan said the main point was that vendors could define boundaries for transferring data across classical systems, with an RDMA-like paradigm that could extend to different accelerator types.
  • Amir followed up on whether QPU measurements would be forwarded over RoCE to a GPU for decoding and asked whether Infleqtion uses standardized libraries such as libfabric or UCX.
  • Ryan described the real-time system as a tightly coupled distributed FPGA/processor network that can behave like a larger composed FPGA system.
  • He said the data path is intended to run largely in FPGA fabric to avoid jitter and processor stalls.
  • Ryan said some implementation components use vendor IP at lower layers, while Infleqtion has developed higher-layer packetization and formatting for dispatching data to GPUs or other compute systems.
  • He mentioned NVIDIA Hololink as a layer that can support RDMA/RoCE-related functionality for the GPU-facing path.
  • Amir clarified that the actual data transfer is FPGA-based; Ryan confirmed that the CPU may initiate or configure operations, but the FPGA fabric should remain in the critical data path.

4. Key Takeaways

  • Neutral-atom dynamic reconfiguration changes the compiler and runtime interface because atom motion and all-to-all connectivity are important execution resources.
  • For long-loop batch workflows, network latency is often less important than allocation and orchestration of quantum and classical resources.
  • For real-time workflows, the open interface problem is harder: users need a way to define classical work inside the QPU loop without binding themselves to one vendor's control stack.
  • MLIR-style heterogeneous kernels may be a useful abstraction for vendor-specific lowering of quantum/classical real-time workflows.
  • Repeat-until-success, mid-circuit measurement, conditional branching, and decoder dispatch need explicit resource accounting if they are to be portable and comparable.
  • FPGA-resident data movement is a practical way to reduce jitter in real-time QPU-to-accelerator workflows, but standardizing that path remains an open systems question.

Files

Clone this wiki locally