Resource assessment of classical and quantum hardware for post-quench dynamics

TL;DR

The paper addresses the problem of evaluating run-time and energy-to-solution for post-quench dynamics on neutral-atom analog QPUs by comparing to classical baselines (MPS and NQS). It combines experimental QPU data with classical simulations, derives scaling laws for both quantum and classical resources, and fits these laws to predict performance on larger systems. The results show that neutral-atom QPUs can be competitive in run-time and energy for medium-duration dynamics, while classical methods face steep scaling with entanglement; the QPU’s advantage grows under bounded-noise and potential error-mitigation strategies. These findings support the viability of energy-efficient analog quantum simulation and point to concrete avenues—such as continuous loading and erasure-based error handling—for extending advantage in scalable quantum materials simulations.

Abstract

We estimate the run-time and energy consumption of simulating non-equilibrium dynamics on neutral atom quantum computers in analog mode, directly comparing their performance to state-of-the-art classical methods, namely Matrix Product States and Neural Quantum States. By collecting both experimental data from a quantum processing unit (QPU) in analog mode and numerical benchmarks, we enable accurate predictions of run-time and energy consumption for large-scale simulations on both QPUs and classical systems through fitting of theoretical scaling laws. Our analysis shows that neutral atom devices are already operating in a competitive regime, achieving comparable or superior performance to classical approaches while consuming significantly less energy. These results demonstrate the potential of analog neutral atom quantum computing for energy-efficient simulation and highlight a viable path toward sustainable computational strategies.

Figures (9)

• Figure 1: Resource estimation for simulating post-quench dynamics of an Ising Hamiltonian, showing both computational time and energy consumption on classical and quantum hardware. The classical data is measured from actual simulations, while the QPU data is estimated based on performance data. The figure highlights the crossover point where classical simulations become less efficient and quantum hardware is expected to offer an advantage.
• Figure 2: (a) Illustration of the different processes required during the rearrangement step. (b) Comparison between the mathematical model (dashed line) describing the probability to have zero defects in the register and experimental data (dots), $N$ being the number of atoms in the register. Here, the error bars are given by the standard deviation of the defect-free distribution over time.
• Figure 3: Run-time of a single 1 ns time step ($dtJ\sim0.01$, in which $J$ is the energy scale of the system) for a quantum quench to the Ising model, measured on an NVIDIA A100 GPU, as a function of system size. Results are shown for both MPS with varying bond dimensions in teal and NQS with different types of networks (containing varying numbers of parameters) in purple. Theoretical scaling fits are overlaid and show good agreement with the data, see App. \ref{['app:fitting']} for details.
• Figure 4: (a) Minimum simulation time required for MPS to achieve convergence as a function of pulse duration $t$ and system size $N$. The black region highlights the pulse durations in which 40Gb of RAM was not enough to achieve convergence. (b) Comparison of total simulation time between classical MPS and estimates for QPU execution across varying system sizes for increasing pulse durations.
• Figure 5: (a) Minimum simulation time required for NQS to achieve convergence as a function of quench durantion and system size. The black region highlights the quench durantions in which simulations do not converge. (b) Comparison of total simulation time between classical NQS and estimates for QPU execution across varying system sizes for increasing pulse durations. Only simulations with $R^2$ below a certain threshold (see text) are considered.