arXiv:2611.00303 · Russian Microelectronics
Application a quantum machine learning algorithm to the applied problem of modeling the thermal regime of a servers
P. V. Buzin, M. A. Remnev
Abstract
Quantum computing is a rapidly developed field of science. In the NISC era, quantum machine learning stands out, potentially outperforming classical machine learning. In this paper, we explore the applicability of quantum machine learning to a real-world application. We solved the problem of prediction of the temperature of GPUs in a high-performance computing cluster. We trained a quantum circuit to predict the temperature of each of the eight GPU of each cluster node based on 16 physical parameters. We demonstrated that quantum machine learning performs on this task just as well as classical machine learning. Furthermore, using model problems, we demonstrated that quantum machine learning can potentially outperform classical machine learning.
AI Summary
The paper investigates whether parameterized quantum circuits can model GPU temperatures from server telemetry and how their regression accuracy compares with classical neural networks. The application maps 16 inputs—power consumption and utilization for eight GPUs—to eight temperature outputs, with deviations from predicted temperatures intended to flag thermal problems. The models combine trainable linear input/output layers with rotation-and-CNOT circuits, optionally using data re-uploading. All quantum computations use noiseless state-vector simulation and exact expectation values. Preliminary experiments compare these circuits with fully connected tanh networks on damped oscillatory, polynomial, and discontinuous piecewise-linear functions, using comparable parameter counts. Re-uploading improves accuracy over the simple circuit; the reported comparisons favor quantum circuits for oscillatory functions and find comparable accuracy for the other examples. For the telemetry application, training uses a random 10% sample from a dataset reported as 270,000 measurements, split 70%/15%/15% into training, validation, and test sets. An eight-qubit, twelve-layer circuit with 496 reported parameters achieves mean absolute temperature error of 1.41°C and maximum error around 15°C. Classical networks achieve 1.46°C with eight neurons per hidden layer and 1.35°C with sixteen; both have maximum errors around 11°C. The practical comparison therefore shows similar accuracy, with the best classical model slightly better. The paper leaves hardware noise, dataset-size dependence, and overfitting for future study, and states that further work is needed to identify conditions under which quantum models could outperform classical approaches.
Error found: The parameter-count formulas and architecture descriptions are inconsistent: the reported quantum and classical counts do not all correspond to the architectures stated in the text or captions, undermining verification of the parameter-matched comparisons. The claimed potential superiority is supported only by comparisons with selected tanh networks on toy functions, not by evidence of an advantage over classical machine learning more broadly. The conclusion extrapolates noiseless, exact-expectation simulation results to real quantum implementation without evaluating finite-shot costs, hardware noise, or execution resources.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Journal reference
- Russian Microelectronics, 2025, Vol. 54, No. 8, pp. 1539-1552
- DOI
- 10.1134/S1063739725601894