A joint study from AWS, NVIDIA, LBNL, and NASA introduces a quantitative model to determine when quantum processors must be physically integrated with supercomputers and when remote cloud access is sufficient for hybrid quantum-classical algorithms
A collaborative research team from Amazon Web Services (AWS), NVIDIA, Lawrence Berkeley National Laboratory (LBNL), and NASA has developed a quantitative framework to assess the infrastructure requirements for integrating quantum processing units (QPUs) with classical high-performance computing (HPC) systems. The model, published as a preprint on arXiv, aims to clarify when quantum hardware must be physically co-located with supercomputers and when standard cloud-based connectivity is adequate for hybrid quantum-classical workflows. This approach is intended to guide infrastructure investment as the field moves toward larger and more reliable quantum systems.
Quantum-Classical Integration Tiers
The study distinguishes between two operational levels in quantum-classical integration. The first, termed the real-time control level, involves sub-microsecond hardware operations such as qubit calibration and quantum error correction (QEC) syndrome decoding. At this level, classical processors must respond within strict latency limits to maintain quantum gate fidelity and prevent logical errors. Physical proximity between QPUs and classical controllers is essential here, as exceeding latency thresholds can cause immediate computational failure.
The second level, the application workflow, covers hybrid algorithms that iteratively exchange data between a classical host and a QPU. Examples include variational quantum eigensolvers and sample-based diagonalization methods. In these cases, classical processing is not constrained by hardware-level timing, so network latency affects total runtime but does not threaten the feasibility of the computation. This distinction allows the model to separate scenarios where tight integration is mandatory from those where remote access is practical.
Benchmarking Communication and Computation
To determine the infrastructure needs for specific algorithms, the researchers introduce a communication-to-computation ratio (Rcc) that quantifies the relative impact of data transfer latency versus processing time. When communication overhead is negligible compared to computation, the workload is compute-bound, and remote cloud access to QPUs is sufficient. Conversely, when data transfer dominates, the workflow is communication-bound, and low-latency physical integration with HPC resources becomes critical for performance.
The model was applied to published benchmarks for two representative hybrid algorithms. For sample-based quantum diagonalization (SQD) on IBM's 77-qubit Heron processor, the classical computation phase was found to dominate the brief quantum measurement step. As a result, the additional latency from standard wide-area network connections had minimal effect on total runtime, supporting the use of remote cloud access for this workload. In contrast, quantum-enhanced Markov chain Monte Carlo (QE-MCMC) algorithms, tested on a 10-qubit IBM device, exhibited high communication overhead due to rapid, repeated exchanges between quantum and classical systems. While this workflow benefits from low-latency connectivity, it does not require the full resources of a supercomputer-only a fast classical controller.
Scaling Toward Fault Tolerance
The study also addresses the transition to fault-tolerant quantum computing, where logical qubits are constructed from many physical qubits using error-correcting codes. As QPU scale increases, the speed of classical decoding and correction becomes a limiting factor for logical clock speed. The model defines crossover points where physical integration with HPC resources becomes necessary, such as when quantum memory or advanced error correction reduces the number of required measurement shots. These thresholds are intended to help the HPC community track hardware progress and plan future infrastructure investments as quantum systems evolve.
According to the authors, the open-source performance model is designed to be updated as new hardware and algorithms emerge. By providing a diagnostic tool grounded in measurable parameters, the framework aims to move infrastructure planning beyond speculation and toward evidence-based decision-making. The full technical details are available in the arXiv preprint and are discussed in further depth on the AWS Quantum Technologies Blog.
For context, the distinction between physical and logical qubits is central to understanding quantum computing infrastructure. Physical qubits are the actual quantum devices-such as superconducting circuits or trapped ions-that can be directly manipulated and measured. Logical qubits, in contrast, are encoded across multiple physical qubits using error-correcting codes to protect against noise and operational errors. Achieving fault tolerance requires not only high-fidelity physical qubits but also fast, reliable classical processing to decode error syndromes and apply corrections in real time. The balance between quantum and classical resources, and the latency of their interaction, sets practical limits on the performance and scalability of quantum computing systems.
