Source-linked AI summary
Reference Architecture of a Quantum-Centric Supercomputer
Seetharami Seelam, Jerry M. Chow, Antonio Córcoles, Sarah Sheldon, Tushar Mittal, Abhinav Kandala, Sean Dague, Ian Hincks, Hiroshi Horii, Blake Johnson, Michael Le, Hani Jamjoom, Jay M. Gambetta
TL;DR
Existing quantum and classical HPC systems operate largely in isolation even though emerging workloads require both, motivating integrated QCSC architectures. The paper presents a reference architecture and three-phase roadmap spanning quantum offload, tightly coupled heterogeneous systems, and deeper co-design, alongside middleware and management mechanisms for hybrid workloads.
Problem
Quantum and classical HPC systems must be integrated because emerging hybrid workloads require both resources, while current systems operate in isolation.
Method
The paper develops a QCSC reference architecture and three-phase roadmap covering infrastructure, software, programming abstractions, applications, resource management, and security.
Results
The architecture organizes progressively deeper integration from quantum offload engines to heterogeneous systems and fully co-designed hybrid computational workflows.
Takeaways & Limitations
QCSC provides a framework for coordinating quantum and classical resources across hybrid workloads with different coupling, scheduling, and data-movement requirements.
Abstract
from arXiv · showhide
Quantum computers have demonstrated utility in simulating quantum systems beyond brute-force classical approaches. As the community builds on these demonstrations to explore using quantum computing for applied research, algorithms and workflows have emerged that require leveraging both quantum computers and classical high-performance computing (HPC) systems to scale applications, especially in chemistry and materials, beyond what either system can simulate alone. Today, these disparate systems operate in isolation, forcing users to manually orchestrate workloads, coordinate job scheduling, and transfer data between systems -- a cumbersome process that hinders productivity and severely limits rapid algorithmic exploration. These challenges motivate the need for flexible and high-performance Quantum-Centric Supercomputing (QCSC) systems that integrate Quantum Processing Units (QPUs), Graphics Processing Units (GPUs), and Central Processing Units (CPUs) to accelerate discovery of such algorithms across applications. These systems will be co-designed across quantum and classical HPC infrastructure, middleware, and application layers to accelerate the adoption of quantum computing for solving critical computational problems. We envision QCSC evolution through three distinct phases: (1) quantum systems as specialized compute offload engines within existing HPC complexes; (2) heterogeneous quantum and classical HPC systems coupled through advanced middleware, enabling seamless execution of hybrid quantum-classical algorithms; and (3) fully co-designed heterogeneous quantum-HPC systems for hybrid computational workflows. This article presents a reference architecture and roadmap for these QCSC systems.
I. INTRODUCTION
Quantum computing is increasingly paired with classical HPC to address quantum-system simulations and application problems that neither paradigm can efficiently handle alone. The paper motivates QCSC workflows across chemistry, materials, optimization, machine learning, and differential-equation applications.
- Quantum and classical HPC are being combined to scale applications, especially in chemistry and materials, beyond what either system can simulate alone.
- QCSC integrates complementary quantum and classical resources for computational tasks involving quantum correlations and broader classical workflows.
- Quantum simulation: Sample-based Quantum Diagonalization uses quantum-generated samples and classical processing to estimate ground-state energies with lower circuit-depth requirements than practical QPE implementations.
- Quantum simulation: Embedding techniques can assign difficult fragments to quantum processors while classical resources preprocess and post-process fragment contributions.
- Broader applications: Hybrid quantum-classical workflows are also being explored for combinatorial optimization, quantum machine learning, and differential-equation simulation.
A. Motivation for Quantum-Centric Supercomputing Systems
QCSC is motivated by current quantum hardware limits and by workloads whose quantum and classical components have different integration requirements. The paper proposes a three-phase roadmap from quantum offload engines to tightly integrated, co-designed quantum-HPC systems.
- Motivation: Current quantum processors face limited qubit counts, gate error rates, reduced connectivity, and execution-speed constraints, motivating integrated classical support and error mitigation.
- Motivation: Hybrid computational problems embed quantum subroutines within larger classical workflows, motivating phased QCSC development.
- Three-phase roadmap: The roadmap progresses from specialized quantum compute engines, through heterogeneous quantum-classical systems, to deeper system co-design.
- Scope: The article presents a reference architecture and multi-year framework spanning infrastructure, platform capabilities, software, programming abstractions, and applications.
- Architecture terminology: Real-time, near-time, and batch-time interactions differ in latency requirements, while temporal and spatial coupling describe coordination and physical-proximity requirements.
II. QCSC APPLICATION PATTERNS THAT GUIDE THE REFERENCE ARCHITECTURE
QCSC application patterns expose different requirements for quantum-classical coordination, ranging from loosely coupled batch workflows to iterative execution. The SQD use case illustrates quantum sampling combined with classical HPC processing and feedback.
- Application patterns: QCSC requirements are grounded in representative hybrid use cases with diverse temporal and spatial coupling needs.
- Use Case 1: In SQD, the QPU generates electronic-configuration samples while classical HPC partitions, recovers, combines, subsamples, and diagonalizes them.
- Use Case 1: The classical workflow performs subspace projection and Hamiltonian diagonalization in a selected determinant basis.
- Use Case 1: The SQD classical process can iterate by feeding diagonalization outputs back into configuration recovery until convergence or an iteration limit.
- Use Case 1: Iterative quantum-classical workflows challenge the loose-coupling model because classical results may need to refine quantum circuit parameters and generate improved samples.
- Use Case 1: SQD uses loose spatial and temporal coupling, with independently managed batch jobs that can run on co-located or cloud-connected systems.
B. Use Case 2: Closed-Loop Electronic Structure Calculations
Closed-loop SQD and related error-mitigation workflows require tighter coordination between quantum hardware and classical HPC. Their repeated data exchange, optimization, and analysis place demands on co-location, latency, bandwidth, and classical compute capacity.
- Closed-loop SQD: Closed-loop SQD iteratively combines quantum sampling, classical subspace diagonalization, and parameter optimization to refine a molecular wavefunction representation.
- Closed-loop SQD: Closed-loop execution requires tight temporal and spatial coupling because each system’s output feeds the next step and must be transferred promptly.
- Closed-loop SQD: Co-location and unified control help dedicate quantum and classical resources, whereas shared cloud access makes coordinated dedication significantly challenging.
- Error mitigation: Error mitigation trades qubit overhead for sampling or runtime overhead and can involve tensor-network calculations, Pauli propagation, and noisy quantum executions.
- Error mitigation: These hybrid workflows extend the practical reach of pre-fault-tolerant quantum systems through parallel classical simulation, data reduction, and mitigation optimization.
- Error mitigation: Advanced error-mitigation workflows shift substantial computational burden to CPUs and GPUs, requiring high-throughput data movement and rapid orchestration.
D. Use Case 4: Open Loop Error Correction Research
Open-loop QEC research uses flexible classical processors connected to quantum systems over high-bandwidth links, while hierarchical correction separates rapid inner stabilization from slower outer noise suppression.
- Open-loop QEC research: QEC research spans hardware, control, coding theory, and classical compute, with dynamic circuits and spacetime checks demanding deterministic timing and high-bandwidth measurement capture.These workloads require tight integration between QPUs and classical resources.
- Open-loop QEC research: Offline decoding supports decoder development, hierarchical partitioning, code comparisons, and threshold estimation without strict real-time feedback.Measurement streams can be sent to GPUs, TPUs, CPUs, or larger HPC resources over high-bandwidth links.
- Hierarchical error correction: Hierarchical error correction uses an inner hardware-adapted code for real-time physical-qubit stabilization and an outer algorithm-adapted code for logical protection and noise suppression.The two coding levels operate at different physical and algorithmic timescales.
- Hierarchical error correction: The inner layer decodes syndrome measurements within ∼10 µs control loops, whereas the outer layer operates on logical qubits at roughly 100 µs to 1 ms.The inner regime uses specialized QPU control electronics, while the outer layer maps naturally to nearby scale-up GPUs or classical accelerators.
- Hierarchical error correction: Low-latency CPU and GPU scale-up systems could enable outer-code research toward fully integrated fault-tolerant systems with hardware-efficient real-time inner correction.This architecture separates rapid inner stabilization from deeper inference and additional noise suppression across logical structures.
F. Summary of the workload patterns for QCSC
QCSC workloads progress from loosely coupled batch jobs to tightly coupled concurrent execution, with increasing demands for spatial proximity, temporal coordination, bandwidth, and low latency.
- Workload patterns: QCSC integration patterns differ in coupling requirements, latency constraints, classical resources, data sovereignty, and security requirements.These dimensions organize the workload patterns summarized in Table I.
- Batch-time integration: The first pattern uses loose batch-time integration, with quantum and classical jobs operating independently without spatial or temporal co-location.Standard SQD workflows exemplify this mode.
- Closed-loop workflows: The second pattern uses tight spatial coupling for closed-loop batch workflows, where one system’s outputs inform the other system’s next iteration.Co-location improves efficiency and keeps job data within organizational boundaries.
- Error mitigation: The third pattern combines multiple noisy quantum circuits with classical tensor-network or Pauli-propagation methods to estimate noise-free results.Future real-time orchestration could improve throughput and reduce quantum-processor idle time during classical post-processing.
- Error correction: Error-correction research requires high-bandwidth communication and specialized GPU or TPU resources, while hierarchical correction requires co-located accelerators, low-latency interconnects, and real-time feedback.These modes impose stronger coupling than independent batch workflows.
- Overall progression: Across the modes, QCSC progresses from weak spatial and temporal coupling through iterative batch loops to strong coupling with concurrent execution and near-time latency.This progression captures the increasing coordination demands of hybrid quantum-classical workloads.
III. QCSC REFERENCE ARCHITECTURE
The QCSC reference architecture integrates quantum systems with heterogeneous HPC resources through layered hardware, orchestration, middleware, and application components, supported by multiple classical proximity tiers and evolving interconnects.
- Architecture overview: The reference architecture has four horizontal layers—Hardware Infrastructure, System Orchestration, Application Middleware, and Applications—plus cloud software, monitoring, and security cross-cutting layers.These layers define the full stack for quantum-HPC integration.
- Quantum System: The Quantum System combines a classical runtime and one or more QPUs connected by a real-time interconnect.Its specialized processors support deterministic, low-latency QPU operations such as QEC decoding, mid-circuit measurement, calibration, and reset.
- Quantum System: A Quantum Systems API abstracts the heterogeneous real-time layer and provides a stable programmatic boundary to orchestration and the rest of the architecture.The interface is intended to be potentially vendor-portable.
- Scale-up systems: Partner scale-up systems place CPU, GPU, and accelerator nodes near quantum systems over low-latency links for tasks such as training and evaluating decoders against syndrome streams.These systems support near-time coordination between quantum and classical resources.
- Scale-out systems: Partner scale-out systems provide flexible cloud or on-premises CPU, GPU, and accelerator resources connected through high-bandwidth interconnects.They support broad classical workloads surrounding and complementing QPU execution and can be tailored to application-specific computation profiles.
- Middleware and workflows: Application middleware supplies parallel programming models for both advanced classical computations and quantum computations, while workflows include preprocessing, post-processing, simulation, and validation.QCSC also requires networking and storage fabrics that bridge quantum-system interfaces with HPC file-system infrastructure.
B. QCSC System Orchestration Layer
The QCSC system orchestration layer coordinates interdependent classical and quantum workloads through unified resource management and workflow execution. Its application middleware extends this coordination to hybrid programming models that combine tensor-based classical computation with quantum circuits and graph-based execution.
- System orchestration: QRMI provides a minimal interface for resource allocation, job scheduling, and job management across quantum and classical resources.It abstracts hardware-specific details for resource managers such as Slurm.
- System orchestration: A Slurm SPANK plugin exposes QPUs as generic resources and invokes QRMI functions to access quantum resources through interfaces such as REST APIs.QRMI can also support vendor backend models and facilitate integration with tools such as Qiskit.
- Workflow management: Workflow managers add explicit execution flow, failure recovery, dependency context, and reproducibility for hybrid quantum-classical workflows.They can launch executables in environments such as MPI or Qiskit, while emerging extensions address multi-tenant QPU use.
- Application middleware: The application middleware layer supports distinct programming models for tensor-based classical computation and quantum-circuit computation.Classical models express tensor dependencies, while quantum algorithms are represented as circuits spanning multiple abstraction levels and hardware-specific instruction sets.
- Application middleware: Tensor Compute Graphs represent hybrid quantum programs as directed acyclic graphs whose nodes can include quantum circuits, parameter-generation graphs, and classical post-processing.This model makes interdependencies and quantum error-mitigation or detection workflows easier to express and orchestrate.
D. QCSC Applications
The QCSC applications layer extends domain-specific solvers with quantum embedding, circuit-synthesis, and application-aware post-processing capabilities. These components decompose workflows across quantum and classical resources while accommodating different classical architectures.
- Quantum embedding: Quantum-embedding capabilities extend traditional embedding methods by encoding domain-specific sub-components for quantum execution.Supported representations include fermionic or bosonic systems for Hamiltonian simulation and graphs for optimization.
- Circuit preparation: Application libraries and solvers translate problem inputs, constraints, and tradeoffs into quantum circuit representations.The preparation pipeline maps domain representations to circuits, optimizes application constraints, and develops system-efficient encoding strategies.
- Circuit preparation: The resulting circuit informs the selection of error-mitigation and error-correction methods and produces the Tensor Compute Graph required for target-hardware execution.This connects circuit preparation with the broader hybrid execution model.
- Post-processing: Quantum libraries also require application-aware post-processing to translate probabilistic quantum outputs into application results.Sample-based Quantum Diagonalization is cited as an example, and this class of algorithms is expected to expand with broader experimentation.
- Classical integration: Classical pre- and post-processing requirements vary with the method and problem encoding, requiring interoperable libraries optimized for different classical architectures.These subroutines enable quantum computation to be integrated into complete application workflows.
E. QCSC Cross Cutting Issues
QCSC integration introduces cross-cutting challenges in orchestration, observability, and resource management as quantum and classical systems become more tightly coupled. Cloud-native HPC trends may provide infrastructure for unified heterogeneous scheduling, while QPU fidelity and scarcity require specialized monitoring and utilization strategies.
- Cloud software: Tightly coupled hybrid workflows expose a mismatch between batch-oriented HPC schedulers and cloud APIs that submit quantum circuits asynchronously.The mismatch becomes a bottleneck when classical and quantum computations must interact iteratively with low latency and co-located resources.
- Cloud software: Growing use of cloud-native stacks, containers, microservices, and Kubernetes in AI-focused HPC may support unified scheduling across CPUs, GPUs, and QPUs.This trend is driven by HPC adoption of AI workloads that include fine-grained services as well as large-scale training.
- Monitoring and resource management: QCSC monitoring must account for quantum-specific factors including fluctuating qubit fidelity and the scarcity of QPU capacity.Effective operation requires current fidelity information for qubit selection and sufficient classical pre- and post-processing to keep QPUs saturated.
- Monitoring and observability: Hybrid quantum-classical systems require observability across the full computational stack to coordinate increasingly integrated workloads.The convergence requires three monitoring developments, including integration of quantum hardware telemetry into HPC observability frameworks.
- Monitoring and observability: Monitoring infrastructure must correlate device-level quantum telemetry with classical nodes, container orchestration, quantum runtimes, schedulers, and hybrid workflows.This cross-layer visibility extends established HPC and AI monitoring practices to heterogeneous QCSC environments.
3) QCSC Security Architecture:
QCSC security must protect heterogeneous quantum-classical programs and data across distributed systems with delay-sensitive execution. The proposed Confidential Code and Data Encapsulation abstraction combines data-access policies, task encapsulation, and a TEE-enhanced runtime, supported by compartmentalization and granular policy enforcement.
- Security requirements: QCSC security challenges arise from heterogeneous programming and data models, cross-system APIs, and inherently delay-sensitive quantum program pathways.The threat model includes protecting provider resources and users’ job code, data, and computation results.
- Security requirements: Confidential execution across distributed heterogeneous QCSC environments requires system-wide automation and programming-model support for data-access policies.TEE configuration and enforcement of complex policies can otherwise be time consuming and error-prone.
- Confidential Code and Data Encapsulation: Confidential Code and Data Encapsulation unifies classical and quantum computation with secure data-access control at the programming level.CDE includes a programming data-access model, task encapsulation module, and TEE-enhanced runtime for policy verification and enforcement.
- Security enforcement: CDE requires mathematically verified task compartmentalization and unified system-wide security governance enforced through granular interposition.Compartmentalization isolates tasks and data across security domains, while policy enforcement points can cover API access and resource operations.
- Security enforcement: QCSC security policies must handle access-control scenarios involving data-center location, quantum-circuit co-location, and cross-talk concerns.Composable security plugins can enforce policies and log or correlate accesses across the system.
IV. QCSC REFERENCE ARCHITECTURE ROADMAP: A PHASED APPROACH
The roadmap progressively integrates quantum systems with classical HPC, beginning with specialized offload engines and advancing toward tightly coupled heterogeneous platforms. Each phase addresses increasing requirements for hardware proximity, resource management, latency, feedback, and hybrid applications.
- Phased roadmap: The roadmap advances through phases that progressively deepen integration between quantum and classical computing systems.The strategy accounts for concurrent advances in classical systems technologies and quantum systems.
- Phase 1: Quantum systems initially serve as specialized compute offload engines within existing HPC infrastructures.This first phase establishes foundational integration across multiple dimensions.
- Phase 1: Early on-premises deployments require careful site preparation and infrastructure planning for reliable hybrid quantum-classical operation.Reported considerations include proximity to classical HPC systems and facility requirements such as vibration minimization.
- Phase 1: Standardized resource interfaces and HPC schedulers are being adapted to discover, allocate, and manage quantum resources within familiar workflows.QRMI and systems such as Slurm lower adoption barriers for existing HPC users.
- Phase 1: As integration matures, applications enable closed-loop hybrid computations and inform requirements for deeper system integration.Scale-up systems co-located with quantum systems also support development of error mitigation and early outer decoder algorithms.
- Phase 2: The second phase couples quantum and classical resources into integrated platforms that reduce latency and support sophisticated feedback mechanisms.Hybrid algorithms use multiple feedback loops across quantum systems, GPUs, specialized accelerators, and scale-out resources.
C. Phase 3: Tightly Integrated Quantum and HPC Systems
The third phase envisions quantum and classical resources as a unified, co-designed platform rather than separate components. This phase is paired with higher-level programming abstractions, multi-tenant execution, and an architecture spanning hardware, resource management, middleware, and applications.
- Phase 3: Phase 3 fully co-designs quantum and classical resources as a unified heterogeneous platform from the ground up.The goal is to architect the systems together rather than integrate standalone components after deployment.
- Phase 3: The architecture follows the broader HPC evolution from PCIe-attached accelerators toward tightly integrated components with optimized interaction.The paper uses GPU integration history as an analogy for quantum-HPC development.
- Programming model: A unified programming model is expected to let developers express complex hybrid algorithms while middleware handles heterogeneous execution complexity.The model is intended to provide higher-level constructs for quantum-classical workflows.
- Phase 3: Multi-tenant quantum execution is expected to emerge alongside the need to address associated security challenges.The paper places this capability within the fully integrated phase.
- Reference architecture: The proposed architecture spans five use-case exemplars, interoperable networking and data fabrics, quantum resource management, workload scheduling, middleware, and applications.These layers support hybrid workloads while leveraging classical and quantum software ecosystems.
- Roadmap evolution: The three-phase architecture is expected to evolve as quantum systems, classical technologies, software, and applications mature.The authors present the roadmap as a reference architecture that the community will refine through adoption and development.