Meeting 2026 07 06

Meeting - July 6, 2026

• Date: Monday, July 6, 2026
• Time: (Slot #1) @ 12 pm - 1 pm US Eastern Daylight Time (View Meeting Time in Your Timezone)
• Location: Virtual Meeting.
• Meeting Link: openqse.org/captcha

Upcoming meetings

• Next: Monday, July 13, 2026 (Slot #2) 7 pm - 8 pm US Eastern Daylight Time (View Meeting Time in Your Timezone)

Agenda

• Architecture Seminar Series: HPE - Namit Anand and Michael Ferguson

Attendance

Person	Institution
Amir Shehata	ORNL
Sakshi Chhabra	IISc
Mahmoud Elkarargy	Brightskies
Laura Schulz	ANL
Michael Ferguson	HPE
Namit Anand	HPE
Abdur Rahman Hatim	IISc
Abdelaziz Rahwan	Brightskies
Eun Kyung Lee	IBM
Doug Oucharek	ORNL
Gabriella Bettonte	E4
John Children	Quantinuum
Nick Wright	NVIDIA
Doriana Medic	University of Turin (UniTO)
Brandon Neth	HPE
Andy Stone	HPE
Oliver Alvarado Rodriguez	HPE
Robert Wisniewski	HPE
Konrad Wojciechowski	PSNC
Oscar Wallis	STFC
Kevin Kissell	Alice & Bob
Thomas Naughton	ORNL
Claudio Siqueira de Carvalho	IBM
Aaron Lott	HPE
Sebastian Stern	Amazon
Lukas Burgholzer	Munich Quantum Software
Patrick Deuley	IonQ
Alain Roy	IonQ
Aurélien Nober	PASQAL
Humberto Munoz Bauza	HPE
Bran Radovanovic	ORNL
Josh Moles	Independent
Neal Erickson	Quantinuum
Andrea Delgado	Qblox
Ahmed Sobhy	Brightskies
Smriti Bajaj	Dell Technologies
Jiri Schindler	Independent
In-Saeng Suh	ORNL
Jeff Heckey	AWS
Roberto Rocco	E4
Elise Jennings	Riverlane
Bartosz Cybulski	Unknown
Yogesh Simmhan	IISc
Katie Klymko	NERSC
Jim Garrison	IBM
Sarah Osama Osman	Brightskies
Ashwin Kumar	Julich
Cathy	Unknown
Ryan Landfield	ORNL
Krzysztof Kurowski	PSNC
Dan Holme	Qoro
Alex Chernoguzov	Quantinuum
Jonathan Skone	NERSC
Tom Beck	ORNL
Valeria Bartsch	FraunhoferCML
Cameron Brunner	Siemens DISW

Executive Summary

• HPE presented an application-centric view of quantum supercomputing, focused on integrating fault-tolerant quantum computing (FTQC) with classical HPC resources.
• The main theme was that useful quantum simulation workflows will often need substantial classical computation before, during, or after the quantum circuit.
• The talk focused primarily on quantum error correction (QEC) and FTQC rather than NISQ systems.
• Namit Anand covered:
  • Why quantum simulation workflows require HPC support.
  • How QEC changes compiler and scheduler priorities.
  • Distributed electronic structure simulation.
  • Entanglement-aware circuit distribution.
  • Initial-state preparation challenges for quantum phase estimation.
  • The role of dynamic and adaptive circuits in reducing state-preparation costs.
• Michael Ferguson covered:
  • Logical mid-circuit classical computation.
  • Adaptive inverse quantum Fourier transform (QFT) followed by measurement.
  • Why approximating fault-tolerant rotation sequences may require helper classical compute outside the control electronics.
  • Externally controlled conditional gates as a system-design mechanism.
  • Quantum oracle sketching, feed-forward QSVT, and possible future feed-forward algorithms.
  • System-level implications for integrating QPUs with co-located HPC nodes over low-latency networks.
• The discussion emphasized latency more than bandwidth for several proposed mid-circuit workflows.
• A key system requirement raised near the end was point-to-point end-to-end latency in the microsecond range rather than milliseconds.

Main Presentation

Framing: Application-Centric Quantum Supercomputing

• Namit framed the talk around accelerating quantum simulation workflows rather than discussing general-purpose computing.
• The motivating application class was Hamiltonian simulation, including:
  • Electronic structure.
  • Materials problems.
  • Fermi-Hubbard-type systems.
  • Phase diagrams and repeated parameter sweeps.
• The representative workflow was quantum phase estimation:
  • Identify the target Hamiltonian.
  • Prepare a good approximate ground state.
  • Run the quantum computation, such as state preparation plus energy estimation.
  • Post-process outputs.
  • Repeat for additional points in the problem space.
• A central point was that the quality of the initial reference state strongly affects the number of phase-estimation repetitions required.
• Preparing a useful reference state can itself require a meaningful classical computation.
• Other algorithms may shift the classical cost elsewhere:
  • Sample-based quantum diagonalization may reduce the need for a highly optimized initial state.
  • But it can require heavy classical post-processing later.
• HPE's proposed quantum supercomputing framework therefore assumes meaningful interaction between QPUs and classical HPC resources.

FTQC and QEC Focus

Why the Talk Focused on Fault Tolerance

• The presenters explicitly focused on QEC and quantum error-corrected computers.
• Some ideas may apply to NISQ systems, but the target workloads require deeper and more reliable circuits than current noisy devices support.
• Namit described the NISQ limitation as a practical depth barrier: after relatively shallow circuits, noise effectively scrambles the useful signal.
• QEC changes the system design problem because:
  • Noisy physical qubits must be redundantly encoded into logical qubits.
  • Some logical gates are not directly available fault-tolerantly.
  • Rotation gates often need approximation, distillation, and injection.
  • Real-time decoding becomes a major classical compute requirement.

Compiler and Scheduler Implications

• In NISQ compilation, minimizing two-qubit gates such as CNOTs is often central because entangling gates are noisy.
• In FTQC, the optimization target changes:
  • T gates and rotation gates become major cost drivers.
  • Magic-state distillation and gate teleportation introduce explicit resource allocation problems.
  • Compilers and schedulers must reason about distillation factories, logical qubit patches, and movement or teleportation of resources to where they are consumed.
• Namit later summarized FTQC scheduling as a workflow where:
  • A job submits a quantum program.
  • Synthesis and FTQC compilation allocate resources.
  • A subset of qubits may be reserved for magic-state distillation.
  • Distilled states are moved or teleported to the computational region.
  • Rotation gates consume those resources when needed.
• This implies schedulers need an intermediate representation that exposes QEC-level resource needs, not only circuit-level gates.

Real-Time Decoding and QLDPC Codes

• Namit highlighted QLDPC codes because they promise better encoding rates.
• Better encoding rates matter because current hardware is still around the order of 100 physical qubits in many platforms.
• Decoder design creates trade-offs among:
  • Accuracy.
  • Latency.
  • Hardware usage.
  • Generality across codes.
• The system should support evolving QEC research:
  • New QEC codes.
  • Different decoder implementations.
  • Solver backends on CPUs, GPUs, FPGAs, or other accelerators.
• The broader design goal is flexibility rather than hard-coding one decoder strategy.

Distributed Electronic Structure Simulation

Distribution Options

• Namit introduced distributed quantum chemistry as the first application instance.
• The central question was how to distribute quantum circuits or Hamiltonian simulation across multiple QPUs.
• Two broad approaches were discussed:
  • Shared entanglement: prepare Bell pairs between QPUs and consume them to implement cross-QPU two-qubit operations.
  • Circuit knitting or adaptive circuit knitting: treat QPUs as independent processing units and stitch results together through classical post-processing.
• Circuit knitting is often described as exponentially costly in the number of cut two-qubit gates.
• Namit argued that this gate-count-centric view can overstate the difficulty for structured Hamiltonian simulation problems.

Entanglement-Aware Cuts

• HPE's goal is not to distribute arbitrary quantum circuits, but to distribute Hamiltonian systems with physical structure and locality.
• For structured Hamiltonians, the overhead should depend on how strongly two subsystems are coupled, not only on how many two-qubit gates are cut.
• The relevant quantity was described as eta, the strength of the boundary Hamiltonian between subsystems.
• The cost should be tied to the boundary interaction and entanglement across the cut.
• Namit emphasized:
  • Uniform systems may not have an obvious best cut.
  • Real systems are heterogeneous because of disorder, defects, and nonuniform structure.
  • Heterogeneous systems have nonuniform spatial entanglement profiles.
  • Low-entanglement regions are better candidates for distribution cuts.
• The high-level message was that finding optimal cuts is really about characterizing entanglement, not just load balancing.

Electronic Structure as a Generalized Lattice

• Electronic structure Hamiltonians were framed as describing electrons hopping among orbitals and repelling each other.
• For distribution purposes, this can be treated as a generalized lattice problem.
• Namit chose Trotterization because it preserves Hamiltonian structure.
• He contrasted this with approaches such as qubitization:
  • Qubitization can have strong asymptotic complexity.
  • But it may hide or wash out locality and other physical structure that is useful for distribution.

Gaussian and Coulomb Layers

• Trotterized electronic-structure time evolution can be viewed as alternating:
  • Gaussian layers.
  • Coulomb layers.
• Gaussian layers were described as special:
  • They are analogous to fermionic Clifford or stabilizer-like operations.
  • Their distribution cost can be characterized efficiently through covariance matrix methods.
  • This avoids gate-by-gate cutting costs that can scale extremely poorly.
• HPE's method was described as improving characterization from an exponential-size object to a linear-size matrix for Gaussian layers.
• Coulomb layers were described as often having order-one distribution cost over relevant time scales.
• The reason given was that disorder in the Coulomb layer leads to localization within fragments.
• Namit connected this to many-body localization:
  • Disorder can make a system behave more like a quasi-integrable model.
  • Entanglement growth and information movement can be constrained.
  • This can yield more efficient descriptions and lower distribution overhead.
• He connected this to the previous week's quantum chemistry discussion, where locality and compact representation also arose as central themes.

Concurrent Compilation

• HPE proposed a workflow with concurrent classical compilation on the HPC side:
  • Simulate or estimate the entanglement structure of the electronic-structure Trotter circuit.
  • Identify low-entanglement cuts dynamically or during preprocessing.
  • Use those cuts to distribute quantum simulation across QPUs.
• The same framework can estimate:
  • Classical overhead for adaptive circuit knitting.
  • Shared-entanglement costs, roughly through the logarithm of the same overhead.
• This supports a unified way to evaluate both classical-networking and quantum-networking distribution strategies.

Resource and Runtime Considerations

FTQC Runtime Estimates

• Namit referenced HPE work on trade-offs between full QEC and partial fault tolerance.
• The numbers discussed in this talk were stated to be full FTQC numbers.
• For smaller fragments, such as roughly 66-orbital problems, quantum runtime could be in the range of seconds to minutes under favorable assumptions.
• The workflow question is whether a large quantum chemistry problem can be divided into smaller fragments, such as 6- to 10-orbital pieces, that can each run on FTQC hardware in a reasonable time.
• The total runtime depends on:
  • Fragment size.
  • Number of logical qubits per fragment.
  • Hardware platform speed.
  • Desired accuracy.
  • Number of classical-quantum iterations.
• Namit noted that fast superconducting-qubit assumptions lead to more optimistic runtimes; slower platforms such as neutral atoms or trapped ions may increase those runtimes substantially.

Classical Compute Requirements

• The required classical compute may include:
  • Initial-state optimization.
  • Entanglement-boundary tracking.
  • Circuit distribution decisions.
  • Sample-based post-processing for some algorithms.
  • Decoder support and QEC experimentation.
• Namit argued that the community may underestimate the cost of initial-state preparation.
• For hard chemistry problems, such as FeMo-cofactor or iron-sulfur clusters, initial-state preparation can be extremely expensive.
• He cited an example where preparing an initial state for a small iron-sulfur-type problem could require as much as 10^10 Toffoli gates.
• This cost can be comparable to, or worse than, the phase-estimation algorithm itself in some settings.
• The broader point was that resource estimates should not hide initial-state preparation under a simplifying assumption.

Communication Timing

• For the distributed quantum simulation workflows discussed by Namit:
  • Communication on the order of seconds could be useful.
  • Waiting hours to days would likely break the intended workflow.
• Later in Michael's part, the latency requirement became much tighter for logical mid-circuit adaptive operations.

Dynamic and Adaptive Circuits

Why Adaptive Circuits Matter

• Namit introduced dynamic circuits as a way to reduce state-preparation costs.
• The focus was on logical-layer adaptivity, not just physical control-layer feed-forward.
• Examples discussed:
  • Long-range CNOTs in superconducting systems can be reduced from order-N depth to order-1 depth using measurements and feed-forward.
  • Preparing tensor network or matrix product states for Fermi-Hubbard or chemistry simulations can receive exponential depth reductions in some settings.
• Namit mentioned reductions from order log N to order log log N for certain tensor-network state preparation cases.
• The underlying structure involved multi-scale entanglement renormalization ansatz ideas.
• The goal is to reduce large state-preparation costs enough to make phase-estimation workflows more practical.

Logical Mid-Circuit Classical Compute

Michael Ferguson's System-Design Claim

• Michael's part focused on classical computation that occurs mid-circuit at the logical algorithm level.
• He distinguished this from QEC-internal real-time decoding or physical control operations.
• The key claim was that some important adaptive logical circuits require classical computations that are too large or too flexible to run inside conventional control electronics.
• Therefore, the quantum resource needs a low-latency path to helper classical compute resources.

Adaptive Inverse QFT Followed by Measurement

• Michael used inverse QFT followed by measurement as the first example.
• A textbook inverse QFT followed by measurement uses order N^2 controlled rotations.
• A known semi-classical procedure can reduce this to order N rotations by:
  • Moving measurements earlier.
  • Replacing quantum-controlled rotations with classically controlled rotations.
  • Combining classically controlled rotations into a single rotation whose angle depends on prior measurement outcomes.
• This is important because inverse QFT followed by measurement appears in many algorithms, including phase-estimation-style algorithms.

Why Control Electronics May Not Be Enough

• In FTQC, arbitrary-angle rotations cannot simply be applied directly.
• A rotation must be approximated by a sequence of allowed fault-tolerant gates.
• In the semi-classical QFT procedure:
  • The required rotation angle depends on previous measurement results.
  • Therefore the approximating sequence must be computed while the circuit is running.
• Michael said current code for approximating such a sequence took about 4 milliseconds on a laptop.
• He argued this is not suitable for simple control electronics in the assumed baseline FTQC system design.
• A helper classical compute resource could:
  • Receive logical measurement outcomes.
  • Compute the required approximating gate sequence.
  • Return the information needed to execute the corresponding conditional gates.

Logical Coherence Timing

• Amir asked whether this interaction must occur within the physical coherence time of the qubits.
• Michael clarified:
  • The discussion is at the logical level.
  • Error correction extends effective logical coherence times substantially.
  • Therefore the timing constraint is not the same as physical-qubit coherence time.
  • But latency still matters because logical errors accumulate over time and because applications slow down if the interaction is too delayed.

Externally Controlled Conditional Gates

Proposed Mechanism

• Michael proposed adding a mechanism to the system model:
  • The quantum resource sends mid-circuit measurement results to helper classical resources.
  • The helper resources send back classical condition values.
  • The quantum circuit contains externally controlled conditional gates.
• This avoids needing to recompile a new gate sequence inside the quantum resource mid-circuit.
• Instead, the circuit can include a preallocated pattern of conditional operations.
• Classical bits from the helper resource determine which operations are actually applied.

Rotation Approximation Strategy

• Michael described using a normal-form representation for single-qubit unitaries.
• A repeating pattern of gates can represent arbitrary single-qubit unitaries.
• The system can allocate enough iterations of the pattern to meet an error bound.
• The helper classical compute sends bits that indicate which gates in the pattern should be active.
• This supports adaptive rotations without a drastic re-architecture of the quantum resource.

Other Mid-Circuit Examples

Quantum Oracle Sketching

• Michael discussed a recent quantum oracle sketching proposal for processing massive classical data.
• The claim of that work, as described in the talk, is a space advantage rather than a time advantage.
• The core operation involves approximating a large data set in a quantum state.
• The circuit applies multi-controlled phase gates:
  • The angle is an algorithm parameter.
  • The controls are determined by sampled classical data.
  • Gates are applied depending on whether a function returns 0 or 1.
• This maps naturally to the same externally controlled conditional-gate mechanism.
• The data sampler may need access to a very large data set or substantial classical memory.
• The connection to the quantum resource may send only small amounts of data per interaction, but latency is critical because many interactions may be needed.

Feed-Forward QSVT and Future Algorithms

• Michael also discussed feed-forward quantum singular value transform (QSVT).
• QSVT is a key building block in optimal Hamiltonian simulation algorithms.
• Feed-forward variants can involve future gate parameters depending on previous measurement results.
• If computing those future gates requires more than simple control-electronics logic, helper classical compute becomes useful.
• Michael suggested similar patterns may appear in future feed-forward quantum machine learning algorithms.
• In such cases, the helper compute may need to do more than gate approximation:
  • It may also run a classical optimizer on a short time scale.

System Design Implications

Hardware Direction

• HPE's proposed direction is quantum system integration with HPC systems.
• The envisioned system includes:
  • One or more quantum resources with FTQC capabilities.
  • Co-located classical HPC nodes.
  • A low-latency high-performance network between them.
  • A variety of classical accelerators where useful, including CPUs, GPUs, FPGAs, or specialized hardware.
• The goal is not just a QPU attached to a job-submission queue.
• The goal is an integrated quantum supercomputer that can support:
  • Classical preprocessing.
  • Classical post-processing.
  • Logical mid-circuit helper compute.
  • QEC decoder experiments.
  • Application-specific accelerators.

Software Direction

• The software stack should extend existing quantum and HPC workflows to support:
  • Mid-circuit logical measurement export.
  • Helper-resource computation.
  • Conditional gate activation controlled by external classical bits.
  • Flexible decoder and QEC experimentation.
  • Integration of specialized hardware into the workflow.
• Michael emphasized building on work already underway rather than requiring a complete redesign.

Latency vs. Bandwidth

• For the approximating-sequence example, Michael estimated:
  • An approximating sequence may be on the order of 300 gates.
  • The data sent back may be around 1,000 bits.
  • Bandwidth is not the main concern.
  • Latency is the main concern.
• For oracle sketching:
  • The data sampler may access a very large data set.
  • The data sent to the quantum resource at each step may still be small.
  • The usefulness of the algorithm depends strongly on how quickly that interaction can occur.
• For the physical layout of an integrated quantum/HPC system:
  • Michael said the QPU and HPC resources do not necessarily need to be in the same rack or room.
  • Different floors or buildings can be acceptable if connected by a high-performance network.
  • The target should be point-to-point end-to-end latency in microseconds rather than milliseconds.

Discussion and Questions

Tensor Networks and Entanglement Tracking

• Andrea asked whether tensor-network methods informed the entanglement-aware distribution approach.
• Namit answered yes.
• He said better classical simulation methods improve the ability to track entanglement boundaries and therefore improve cut placement.
• For electronic structure, many-body localization can induce tensor-network-like structure.
• Namit said the team expected a paper soon on the structure they observed.

Trotter Order

• Andrea asked whether first-order Trotterization was used only for a toy problem or whether it was sufficient.
• Namit said higher-order Trotter formulas, such as second order, could be used.
• First-order Trotterization was chosen to expose the Gaussian/Coulomb structure cleanly.
• With sufficiently small time steps, the first-order structure was adequate for the points being made.
• Second-order Trotterization would preserve a similar structure, essentially adding forward and backward versions of the same layers.

Throughput for Mid-Circuit Measurements

• Andrea asked whether the team had throughput estimates, especially for mid-circuit measurement on many qubits.
• Michael did not have concrete throughput numbers.
• He noted that FTQC already requires frequent measurement of many physical qubits for error correction, especially ancilla qubits.
• His assumption was that the quantum resource must already handle heavy internal measurement traffic.
• At the logical layer, additional logical measurements should be modest compared with the QEC-internal measurement stream.

Data Transfer for Approximating Sequences

• Amir asked how much data would be transferred for approximating sequences as problem size grows.
• Michael said the data volume is small:
  • Approximation sequences may involve around 300 gates.
  • The return data might be around 1,000 bits.
  • The measurement data are also small, often bit-level.
• Again, the issue is latency rather than bandwidth.

Data Sampler Freshness

• Jiri asked how fresh the oracle-sketching data needs to be.
• Michael said the answer likely depends on the application.
• The data might be:
  • Loaded from a large static data set in memory.
  • Produced by some computation during the workflow.
• More stringent timing requirements would apply if data freshness depends on the same running circuit or a tightly coupled computation.
• Michael said this is still an open question and depends on how the oracle-sketching work is applied.

Physical Deployment Distance

• Amir asked whether the quantum and HPC resources need to be physically close.
• Michael said high-performance networks can span floors or buildings.
• The important factor is the achieved latency.
• He suggested designing for microsecond-scale point-to-point latency rather than millisecond-scale latency.

FTQC Scheduling Requirements

• Amir asked Namit to elaborate on FTQC scheduling requirements.
• Namit explained that rotation gates require special handling in FTQC.
• A QPU may need explicitly allocated regions for magic-state distillation.
• Distilled states are transported or teleported to the computational region and consumed when rotation gates are applied.
• This means FTQC scheduling must reason about:
  • Distillation factories.
  • Logical qubit allocation.
  • State production rates.
  • Movement or teleportation of resources.
  • Where and when gates are consumed.

Conclusions

• The meeting presented a system-level argument that practical quantum simulation will require integrated quantum and HPC resources.
• The classical resource is not merely for offline preprocessing:
  • It may participate during compilation.
  • It may track entanglement boundaries.
  • It may compute mid-circuit gate approximations.
  • It may run application-specific data samplers or optimizers.
  • It may support QEC decoding and decoder experimentation.
• Entanglement-aware distribution can reduce overhead compared with naive gate-by-gate circuit cutting.
• Initial-state preparation is a major resource-estimation issue and should be accounted for explicitly.
• Dynamic and adaptive logical circuits can reduce state-preparation and algorithmic costs, but they create new system requirements.
• For some workflows, bandwidth is modest but latency is critical.
• HPE's design target is a flexible quantum supercomputer architecture with QPU resources, HPC nodes, low-latency networking, and support for specialized classical hardware.

Open Questions and Follow-Ups

• No formal action items were assigned during the meeting.
• Useful follow-up topics raised by the discussion:
  • Quantify throughput requirements for logical and physical measurement paths.
  • Develop concrete latency and bandwidth targets for mid-circuit helper-compute workflows.
  • Separate requirements for static data sampling vs. data generated during the same computation.
  • Explore tensor-network methods for entanglement tracking in distributed electronic structure simulation.
  • Compare first-order and higher-order Trotter formulas for cut placement and entanglement profiles.
  • Specify FTQC scheduler abstractions for distillation factories, logical resource movement, and gate consumption.
  • Identify which adaptive algorithms can run entirely in control electronics and which require helper HPC resources.
  • Define benchmark workloads that exercise preprocessing, mid-circuit helper compute, and post-processing in one integrated quantum/HPC system.

Files

• Download Transcript