HPC, GPU, and AI Co-Scheduling — Operating Quantum Computers

State-of-the-art quantum systems are hybrid systems. The useful computer is not the QPU alone; it is a coordinated service that combines QPU access, CPUs, GPUs, storage, controllers, simulators, decoders, optimizers, and workflow managers.

Qiskit Serverless, Amazon Braket Hybrid Jobs, CUDA-Q, NVQLink, and cudaq-realtime all point in the same direction: production quantum computing is a heterogeneous scheduling problem. [R81] [R97] [R99] [R105]

55.1 Heterogeneous resource graph

Diagram loads as you read

55.1 Heterogeneous resource graph · Figure 1

View diagram source

flowchart TB
    Workload[Hybrid workload] --> CPU[CPU preprocessing]
    Workload --> GPU[GPU simulation / decoding / ML]
    Workload --> QPU[QPU execution]
    Workload --> Store[Artifact storage]
    Workload --> Net[Low-latency network]
    CPU --> Orchestrator[Workflow orchestrator]
    GPU --> Orchestrator
    QPU --> Orchestrator
    Store --> Orchestrator
    Net --> Orchestrator

The scheduler must reason about resource coupling. A QPU reservation is wasted if the optimizer, decoder, or artifact pipeline is not ready.

55.2 Workload classes

Class	QPU dependency	Classical dependency	Scheduling concern
parameter sweep	repeated QPU tasks	moderate CPU	batching and queue cost
variational loop	iterative QPU calls	optimizer state	session continuity
chemistry estimation	many observables	tensor/data preprocessing	artifact volume
error-correction test	real-time syndrome stream	GPU decoder	latency and throughput
digital twin	no QPU required until validation	high GPU/CPU	simulation capacity
calibration optimization	QPU plus controller	Bayesian/ML loop	safety and rollback

Diagram loads as you read

55.2 Workload classes · Figure 2

View diagram source

flowchart LR
    Workloads[Workload classes] --> Sweep[Parameter sweep]
    Workloads --> VQA[Variational loop]
    Workloads --> Chemistry[Chemistry estimation]
    Workloads --> QEC[QEC test]
    Workloads --> Twin[Digital twin]
    Workloads --> Cal[Calibration optimization]

55.3 Co-scheduling loop

Diagram loads as you read

55.3 Co-scheduling loop · Figure 3

View diagram source

sequenceDiagram
    participant Planner as Planner
    participant CPU as CPU pool
    participant GPU as GPU pool
    participant QPU as QPU scheduler
    participant Store as Artifact store

    Planner->>CPU: Prepare circuits/data
    Planner->>GPU: Simulate or pre-optimize
    Planner->>QPU: Reserve/submit quantum tasks
    QPU-->>Planner: Execution window
    Planner->>GPU: Allocate decoder/optimizer window
    QPU->>Store: Write raw outputs
    GPU->>Store: Write post-processing outputs
    Planner-->>Planner: Decide next iteration

55.4 Capacity reservation

A hybrid job needs a bundle, not a single queue slot.

Illustrative listing · yaml

resource_bundle:
  job_id: vqa-session-204
  qpu:
    backend: target-qpu
    window: 2026-04-19T13:00Z/2026-04-19T15:00Z
    max_shots: 500000
  cpu:
    cores: 64
    memory_gb: 256
  gpu:
    type: h100-or-equivalent
    count: 2
    latency_class: standard | realtime
  storage:
    raw_output_gb: 500
    retention_policy: governed
  network:
    realtime_required: false

QPU-first scheduling underestimates production complexity. Resource-bundle scheduling is the safer abstraction.

55.5 Real-time island

Some workflows require a real-time island close to the controller.

Diagram loads as you read

55.5 Real-time island · Figure 4

View diagram source

flowchart LR
    QPU[QPU] <--> Controller[FPGA/control electronics]
    Controller <--> NIC[Low-latency NIC]
    NIC <--> GPU[GPU decoder/feedback]
    GPU --> SlowCloud[Slow cloud workflow]
    SlowCloud --> Planner[Planner and artifact system]

Do not route microsecond feedback through services designed for human-scale latency.

55.6 AI co-pilots in operations

AI can help tune calibrations, classify anomalies, route workloads, and triage incidents. It should not silently control a QPU without bounded authority.

Diagram loads as you read

55.6 AI co-pilots in operations · Figure 5

View diagram source

flowchart TB
    Telemetry[Telemetry stream] --> Model[AI model]
    Model --> Suggest[Suggested action]
    Suggest --> Guardrail[Policy guardrail]
    Guardrail --> Human{Requires human approval?}
    Human -- yes --> Review[Operator review]
    Human -- no --> Execute[Automated action]
    Execute --> Audit[Audit log]
    Review --> Audit

55.7 Scheduler objectives

Schedulers need multiple objectives.

Diagram loads as you read

55.7 Scheduler objectives · Figure 6

View diagram source

flowchart TD
    Objective[Scheduler objective] --> Fidelity[Maximize expected result quality]
    Objective --> Cost[Control cost]
    Objective --> Fairness[Provide tenant fairness]
    Objective --> Utilization[Use scarce QPU capacity]
    Objective --> Evidence[Preserve evidence completeness]
    Objective --> Latency[Meet latency class]

A job that maximizes utilization can harm science if it runs during poor calibration state. A job that maximizes fidelity can waste scarce capacity if it ignores cost.

55.8 Admission algorithm sketch

Illustrative listing · pseudo

function admit_hybrid_job(job):
    if not evidence_policy_satisfied(job):
        return rejected(["EVIDENCE_POLICY_UNSATISFIED"])
    estimate = estimate_resources(job)
    targets = find_targets(job.required_features)
    for target in ranked(targets):
        # Acquire the whole bundle atomically, or release any partial holds.
        bundle = try_reserve_bundle(target, estimate.cpu, estimate.gpu, estimate.storage)
        if bundle is none:
            continue
        if bundle.satisfies(job.slo):
            return accepted(bundle)
        release_bundle(bundle)
    return rejected(["NO_FEASIBLE_RESOURCE_BUNDLE"])

55.9 Operations metrics

Metric	Why it matters
qpu-window utilization	measures scarce resource use
bundle wait time	captures hybrid scheduling pain
preflight failure rate	shows workflow maturity
GPU decoder headroom	protects real-time QEC workflows
artifact write latency	prevents result loss and backpressure
cost per useful precision	ties resource use to scientific output

Diagram loads as you read

55.9 Operations metrics · Figure 7

View diagram source

flowchart LR
    Metrics[Co-scheduling metrics] --> Util[QPU-window utilization]
    Metrics --> Bundle[Bundle wait time]
    Metrics --> Preflight[Preflight failures]
    Metrics --> Decoder[Decoder headroom]
    Metrics --> Artifacts[Artifact latency]
    Metrics --> Cost[Cost per useful precision]

The future quantum data center is neither a classical data center with a QPU attached nor a physics lab with a scheduler bolted on. It is a heterogeneous operating environment.