Skip to content

Meeting 2026 07 06

amirshehataornl edited this page Jul 9, 2026 · 7 revisions

Meeting - July 6, 2026

Upcoming meetings

Agenda

  • Architecture Seminar Series: HPE - Masoud Mohseni

Attendance

Person Institution
Amir Shehata Oak Ridge National Laboratory
Sakshi Chhabra Indian Institute of Science
Mahmoud Elkarargy Brightskies
Laura Schulz Argonne National Laboratory
Michael Ferguson Hewlett Packard Enterprise
Namit Anand Hewlett Packard Enterprise
Abdur Rahman Hatim Indian Institute of Science
Abdelaziz Rahwan Brightskies
EUN KYUNG LEE IBM
Doug Oucharek ORNL
Gabriella Bettonte E4
John Children Quantinuum
Nick Wright NVIDIA
Doriana Medic University of Turin
Brandon Neth Hewlett Packard Enterprise
Andy Stone Hewlett Packard Enterprise
Oliver Alvarado Rodriguez Hewlett Packard Enterprise
Robert Wisniewski Hewlett Packard Enterprise
Konrad Wojciechowski Poznan Supercomputing and Networking Center
Oscar Wallis STFC
Kevin K. A&B
Thomas Naughton III Oak Ridge National Laboratory
Claudio Siqueira de Carvalho IBM
Aaron Lott Hewlett Packard Enterprise
Sebastian Stern Amazon
Lukas Burgholzer Munich Quantum Software Company
Patrick Deuley IonQ
Alain Roy IonQ
Aurélien NOBER Pasqal
Humberto Munoz Bauza Hewlett Packard Enterprise
Bran Radovanovic Oak Ridge National Laboratory
Josh Moles Independent
Neal Erickson Quantinuum
Andrea Delgado QBlox
Ahmed Sobhy Brightskies
Smriti Bajaj Dell
Jiri Schindler Independent
In-Saeng Suh Oak Ridge National Laboratory
Jeffrey Heckey Amazon
Roberto Rocco E4
Elise Jennings Riverlane
Bartosz Cybulski
Yogesh Simmhan Indian Institute of Science
Katie Klymko NERSC
Jim Garrison IBM
Sarah Osama Osman Brightskies
Ashwin Kumar Julich
SCH 204 Meeting Room
Cathy
Ryan Landfield Oak Ridge National Laboratory
Krzysztof Kurowski Poznan Supercomputing and Networking Center
Dan Holme Qoro
Alex Chernoguzov Quantinuum
Jonathan Skone NERSC
Tom Beck Oak Ridge National Laboratory
Valeria Bartsch Fraunhofer CML
Cameron Brunner Siemens

Notes

  • Namit Anand
  • Michael Paul Pratt Ferguson

Meeting Objective

HPE presented their Application-Centric Quantum Supercomputing Framework, focusing on how future fault-tolerant quantum computers (FTQC) should be tightly integrated with HPC systems. The presentation emphasized:

  • Distributed quantum simulation
  • Quantum error correction (QEC)
  • Classical compute requirements for FTQC
  • Mid-circuit classical computation
  • System architecture for integrated Quantum + HPC environments

Presentation Part 1 – Namit Anand

1. Motivation

The presentation focused on accelerating quantum simulation, particularly:

  • Electronic structure simulation
  • Materials science
  • Chemistry
  • Fermi-Hubbard models

Rather than treating quantum computers as standalone accelerators, HPE proposes viewing them as part of a larger HPC workflow.

Typical workflow:

  1. Classical preprocessing
  2. Initial state preparation
  3. Quantum computation
  4. Classical post-processing
  5. Repeat over parameter space

A key message was that classical computation remains an essential component, even in a fault-tolerant quantum computing era.


2. Importance of Initial State Preparation

Quantum Phase Estimation (QPE) requires a high-quality initial state.

Observations:

  • Better initial states reduce the number of QPE iterations.
  • Preparing these states is itself computationally difficult.
  • A dedicated HPC cluster may be needed solely to prepare high-quality initial states before execution on the quantum computer.

Key point:

Initial state preparation is significantly underestimated in most resource estimates.


3. Focus on Fault-Tolerant Quantum Computing (FTQC)

The presentation intentionally focused on:

  • Quantum Error Correction (QEC)
  • Logical qubits
  • Fault-tolerant algorithms

rather than NISQ systems.

Major implications include:

  • Shift from minimizing CNOT gates to minimizing T gates.
  • Rotation gates become expensive due to magic state distillation.
  • Increased compiler complexity.
  • Significant real-time decoding requirements.

4. Quantum Error Correction Challenges

Discussion centered on QLDPC codes.

Trade-offs identified:

  • Accuracy
  • Latency
  • Hardware utilization

No decoder simultaneously optimizes all three.

Consequences:

  • Different workloads require different decoding strategies.
  • System architecture must support evolving QEC algorithms.
  • CPU, GPU, FPGA, and future accelerators should all be supported.

5. Distributed Quantum Chemistry

Two approaches to distributing quantum circuits were discussed.

Shared Entanglement

  • Bell pairs connect QPUs.
  • Two-qubit gates executed across devices.

Circuit Knitting

  • Separate QPUs operate independently.
  • Classical post-processing reconstructs the full computation.

Traditional circuit cutting incurs exponential overhead.

HPE proposes reducing this overhead by exploiting physical properties of the simulated system.


6. Physics-Based Distribution Strategy

Rather than distributing circuits evenly, HPE recommends distributing based on:

  • Entanglement structure
  • Hamiltonian locality

Key insight:

Distribution cost depends on:

  • boundary entanglement

rather than simply:

  • number of cut gates

This allows substantially lower communication costs.


7. Electronic Structure Simulation

Electronic structure Hamiltonians naturally decompose into:

  • Gaussian layers
  • Coulomb layers

Important findings:

Gaussian Layers

Using covariance matrix techniques:

  • Exponential improvement in estimating distribution cost
  • Linear-sized representations instead of exponential

Coulomb Layers

Observed behavior similar to:

  • Many-body localization

Consequences:

  • Low entanglement
  • Near constant distribution cost
  • Efficient fragmentation

8. Benefits of Application-Aware Compilation

Instead of generic quantum compilers:

The framework continuously:

  • Simulates entanglement
  • Finds optimal cut locations
  • Minimizes communication
  • Estimates Bell-pair requirements

Result:

Orders-of-magnitude reduction in distribution complexity.


9. Resource Estimates

Examples showed that relatively small FTQC fragments (approximately 6–10 orbital problems) could complete in:

  • seconds
  • minutes
  • occasionally hours

Larger problems could be decomposed into these fragments.

Runtime depends on:

  • number of logical qubits
  • communication
  • hardware technology

10. Classical Compute Requirements

Classical resources are required for:

  • Initial state optimization
  • Entanglement tracking
  • Distribution planning
  • Post-processing

HPE argued that future quantum supercomputers will require dedicated HPC resources operating alongside the quantum processor.


11. Initial State Preparation Cost

A major emphasis was placed on this issue.

For difficult chemistry problems:

Initial state preparation may require:

  • ~10¹⁰ Toffoli gates

In some cases:

State preparation may cost more than phase estimation itself.

This cost is often omitted from published resource estimates.


12. Dynamic Circuits

HPE proposed using:

  • measurements
  • feed-forward
  • logical adaptive circuits

Benefits include:

  • Faster long-range operations
  • Shallower circuits
  • Significant reductions in state preparation depth

Discussion

Tensor Networks

Question (Andrea):

Was tensor-network theory used to determine optimal circuit partitions?

Response

Yes.

Tensor-network methods:

  • track entanglement
  • identify optimal cut locations
  • improve classical simulations
  • naturally emerge from many-body localization.

A forthcoming paper was mentioned.


Trotterization

Andrea asked why first-order Trotterization was used.

Response:

  • Mainly to expose Gaussian/Coulomb structure.
  • Second-order methods could also be used.
  • Small time steps preserve accuracy while maintaining the desired decomposition.

Presentation Part 2 – Michael Ferguson


13. Mid-Circuit Classical Computation

Focus shifted to interactions between:

  • Quantum hardware
  • External classical resources

Primary example:

Quantum Fourier Transform (QFT).


14. Quantum Fourier Transform Optimization

Standard QFT:

O(N²) rotation gates.

Using adaptive circuits:

Reduced to:

O(N)

This is the well-known semi-classical QFT.


15. Limitation of Existing FTQC Architectures

The adaptive version requires:

  • computing rotation angles
  • synthesizing gate approximations
  • during execution

Current gate synthesis takes approximately:

4 ms on a laptop.

This is too expensive for today's control electronics.


Key Architectural Insight

Current assumptions are insufficient.

Future FTQC systems require:

Quantum Resource ⇄ HPC Compute

during program execution.


16. Externally Controlled Conditional Gates

Rather than recompiling circuits during execution:

HPE proposes:

  • externally controlled conditional gates

Workflow:

  1. Logical measurement
  2. Send measurement to HPC
  3. HPC computes gate sequence
  4. Send classical control bits back
  5. Continue execution

Minimal data transfer is required.

Latency—not bandwidth—is the limiting factor.


17. Oracle Sketch

Michael discussed the recent Quantum Oracle Sketch paper.

Purpose:

Represent massive classical datasets using exponentially less quantum memory.

Requirements:

  • External data sampler
  • Large classical memory
  • Low-latency communication

Again:

HPC integration becomes essential.


18. Future Adaptive Algorithms

HPE expects similar requirements for:

  • Feed-forward QSVT
  • Quantum machine learning
  • Future adaptive algorithms

These algorithms may require:

  • optimization
  • parameter computation
  • classical inference

during quantum execution.


19. Latency Requirements

Estimated FTQC clock rates were compared with network latency.

Observation:

Modern HPC interconnects operate on compatible timescales.

Microsecond-scale communication is acceptable.

Millisecond or scheduler-level delays are not.


20. Proposed Quantum Supercomputer Architecture

Recommended architecture:

  • Quantum processing units
  • Co-located HPC nodes
  • CPUs
  • GPUs
  • FPGAs
  • Specialized accelerators
  • High-performance, low-latency interconnect

Goals:

  • Flexible experimentation
  • Mid-circuit classical compute
  • Quantum error correction research
  • Support for custom hardware accelerators

Q&A Summary

Andrea

Throughput vs Latency

Question:

Is throughput more important than latency for control systems?

Response:

Large numbers of measurements already occur continuously for QEC.

Logical-level measurements add relatively little additional burden.

Precise throughput estimates are not yet available.


Amir

Data Transfer Requirements

Question:

How much data must be transferred for gate approximation?

Response:

Only approximately:

  • hundreds to thousands of bits

Latency matters much more than bandwidth.


Oracle Sketch Data Flow

Question:

What are the communication requirements?

Response:

Only sampled data is transmitted.

Fast communication is necessary to keep the quantum processor productive.


Physical Placement

Question:

Must HPC and quantum hardware be physically adjacent?

Response:

No.

Different floors or buildings are acceptable.

Requirement:

Microsecond-scale HPC networking.


Jiri

Freshness of Sampled Data

Question:

Does Oracle Sketch require continuously updated data?

Response:

Depends on the application.

Some workloads may sample static datasets.

Others may require computationally generated data.


Amir (Closing Question)

Question:

How does FTQC change scheduling?

Response:

Future compilers must schedule:

  • Magic state factories
  • Distillation resources
  • Logical qubit allocation
  • Gate teleportation
  • Resource movement across the processor

Scheduling becomes significantly more dynamic than current quantum compilers.


Key Takeaways

  • Future quantum computing will rely heavily on HPC integration rather than standalone QPUs.
  • Classical computation remains essential for preprocessing, state preparation, decoding, compilation, and adaptive execution.
  • Initial state preparation is a major, often overlooked, computational bottleneck.
  • Entanglement-aware circuit partitioning substantially reduces distributed quantum simulation costs.
  • Mid-circuit classical computation requires low-latency HPC resources beyond traditional quantum control electronics.
  • Network latency is a primary system design concern; bandwidth requirements are relatively modest.
  • HPE advocates a tightly integrated quantum-HPC architecture with flexible classical resources connected via high-performance networking.

Action Items / Follow-Up

Item Owner
Share upcoming paper on tensor-network structures for electronic structure simulation HPE (Namit)
Continue discussion on FTQC scheduling and compiler architecture in a future session HPE / ORNL
Explore latency and throughput requirements for logical-level adaptive circuits HPE & ORNL
Investigate hardware requirements for externally controlled conditional gates HPE
Continue collaboration on integrated Quantum-HPC system architecture openQSE participants

Files

Clone this wiki locally