A cross-institutional research team comprising scientists from Amazon Web Services (AWS), NVIDIA, Lawrence Berkeley National Laboratory (LBNL), and NASA has published a quantitative performance model to evaluate when quantum processing units (QPUs) must be tightly co-located with classical high-performance computing (HPC) supercomputers—and when standard remote cloud connectivity is fully sufficient.
Published on arXiv and highlighted on the AWS Quantum Technologies Blog, the study cuts through infrastructure speculation to provide a diagnostic tool that guides capital expenditure decisions as the industry transitions toward fault-tolerant systems.
Differentiating Real-Time Hardware Control from Application Workflows
A central contribution of the paper is explicitly separating quantum-classical connectivity into two distinct operational tiers:
- The Real-Time Control Level: Focuses on sub-microsecond hardware tasks, including physical qubit calibration and quantum error correction (QEC) syndrome decoding. Here, classical decoders must process error syndromes and apply corrections within strict physical limits. At this level, low-latency, tightly coupled physical co-location is non-negotiable; exceeding latency thresholds leads to immediate logical gate failure.
- The Application Level: Evaluates hybrid algorithms (such as variational solvers and sample-based diagonalizations) that iteratively exchange data between a classical host and a QPU. At this level, classical processing does not need to complete within sub-microsecond hardware limits, meaning network latency impacts total runtime performance rather than physical feasibility.
Evaluating Infrastructure Regimes: Compute-Bound vs. Communication-Bound
To determine whether a specific hybrid algorithm requires physical co-location with an HPC supercomputer, the authors decompose the total execution cycle into classical compute time, quantum compute time, and communication overhead. This framework establishes a diagnostic communication-to-computation ratio (Rcc) that defines two distinct operational regimes:
- Compute-Bound Workloads: Communication overhead is negligible relative to execution time. Co-locating the QPU next to a supercomputer offers no measurable performance gain, making standard cloud-based remote access fully sufficient.
- Communication-Bound Workloads: The workflow is bottlenecked by data transfer latency. Tight, low-latency physical integration materially improves total execution speed.
Benchmarking Near-Term Workloads and Future Fault Tolerance
Applying the model to published algorithmic benchmarks yielded concrete infrastructure conclusions:
- Sample-Based Quantum Diagonalization (SQD): Evaluated on IBM’s 77-qubit Heron processor for quantum chemistry, SQD demonstrated that because the heavy classical compute phase dominates the brief quantum measurement, communication overhead over standard wide-area networks is negligible—proving that remote cloud access works effectively.
- Quantum-Enhanced Markov Chain Monte Carlo (QE-MCMC): Benchmarked on a 10-qubit IBM device, QE-MCMC demonstrated high communication overhead under remote access due to rapid, repeated round-trips. However, while this workflow benefits from low-latency connectivity, the classical step requires minimal processing power, meaning it requires a low-latency connection rather than a full supercomputer.
- Fault-Tolerant Scaling: As QPUs scale to logical qubits, low-latency integration between QPUs and classical HPC resources becomes increasingly vital, as reaction times directly set the logical clock speed of fault-tolerant execution.
By defining specific crossover conditions—such as shot-budget reductions via amplitude estimation or the emergence of practical quantum memory—the open-source performance model provides a living framework for the HPC community to track hardware progress and plan infrastructure investments.
Review the complete theoretical paper on arXiv Physics here, and read the joint architectural analysis on the AWS Quantum Technologies Blog here.
July 20, 2026
