Researchers from Amazon Web Services (AWS), Nvidia, Lawrence Berkeley National Laboratory, and NASA proposed a model to evaluate the interaction between quantum processors and classical supercomputers.
This model helps determine when quantum hardware should be located near high-performance computing infrastructure and when remote access via the cloud is sufficient.
Requirements Depend on Task Type
The researchers categorized quantum-classical interaction into two levels:
- The first involves real-time equipment management, including the calibration of physical qubits and the processing of quantum error correction data.
- The second encompasses hybrid algorithms, where classical computers and quantum processors alternately perform parts of the computation.
In error correction, the classical system analyzes control measurements to identify which errors need fixing. This processing must occur within intervals ranging from fractions of microseconds to several microseconds. For such tasks, a direct connection with minimal latency is essential; remote access through a standard network is inadequate.
In the second case, latency affects the duration of the operation but does not always determine whether the algorithm can be executed. Therefore, some tasks can rely on cloud access to the quantum device.
Remote Access Isn't Suitable for All Algorithms
The authors divided the execution time of a hybrid algorithm into classical computations, quantum processor operations, and data exchanges.
If the majority of the cycle involves information processing, reducing network latency does little to speed up the computation. However, if the algorithm requires frequent exchanges, a fast connection becomes critical.
The first example was quantum diagonalization based on sampling (SQD), used in quantum chemistry. The authors utilized data from an experiment with 77 qubits on the IBM Heron processor. In this scenario, the bulk of the time was spent on classical processing of results, with communication costs accounting for about 0.0001, making remote access to the quantum processor sufficient.
In contrast, the Markov chain Monte Carlo method with quantum acceleration (QE-MCMC), tested on a 10-qubit IBM device, required frequent sequential exchanges. With remote access, communication costs exceeded computation time by approximately 1000 times.
However, it did not require the power of a full supercomputer; a small classical controller with low-latency connection was adequate.
Error Correction Will Increase Requirements
As the field moves towards logical qubits and fault-tolerant quantum computers, the role of classical infrastructure will grow. The system must continuously process measurements, detect errors, and send corrective commands. The speed of the classical component will begin to dictate the pace of the quantum computer's operation.
The authors anticipate that large fault-tolerant installations will require closer integration with high-performance computing systems. There is no one-size-fits-all architecture; requirements will vary based on the algorithm, hardware characteristics, number of exchanges, and volume of data transmitted.
The researchers suggest using the model for regular reassessment of infrastructure as quantum processors and computational methods evolve.
In July, Nvidia released the code for an AI module for preprocessing quantum error signals. In simulations, coupling with a classical decoder reduced the frequency of logical failures and accelerated processing.