QAOA in Quantum Datacenters: Parallelization, Simulation, and Orchestration

TL;DR

The work tackles the scalability gap in quantum computing by presenting an automated, cloud-based QAOA workflow that partitions large problems, generates batch simulations, and orchestrates distributed execution. It introduces Divi for automated problem decomposition and batch generation, and Maestro as a unified interface to multiple simulators with optimized multi-shot and distributed execution, complemented by a cloud platform for resource management. Empirical results show that partitioned QAOA can match classical baselines on small graphs and improve with larger subproblems, while automated simulation backends and cloud orchestration dramatically reduce end-to-end runtimes. Collectively, the framework lowers infrastructure barriers, accelerates quantum program development, and lays the groundwork for scalable quantum datacenters with hardware-accelerated and distributed simulation capabilities. The approach is poised to influence practical quantum advantage by standardizing automated workflows and enabling seamless integration of heterogeneous resources, including potential future extensions to noise modeling and GPU-accelerated simulators.

Abstract

Scaling quantum computing requires networked systems, leveraging HPC for distributed simulation now and quantum networks in the future. Quantum datacenters will be the primary access point for users, but current approaches demand extensive manual decisions and hardware expertise. Tasks like algorithm partitioning, job batching, and resource allocation divert focus from quantum program development. We present a massively parallelized, automated QAOA workflow that integrates problem decomposition, batch job generation, and high-performance simulation. Our framework automates simulator selection, optimizes execution across distributed, heterogeneous resources, and provides a cloud-based infrastructure, enhancing usability and accelerating quantum program development. We find that QAOA partitioning does not significantly degrade optimization performance and often outperforms classical solvers. We introduce our software components -- Divi, Maestro, and our cloud platform -- demonstrating ease of use and superior performance over existing methods.

Figures (7)

• Figure 1: The parallelization flow used for QAOA with VQE. The flow begins with a principle graph which gets partitioned. From there, individual optimization programs are executed using VQE for finding the ground state of $H_i$. The optimal parameters are used to recover the optimal solution of each subproblem, which aggregate to form the final solution.
• Figure 2: Partitioned QAOA evaluated in comparison to the Goemans-Williamson solver for varying sized problems.
• Figure 3: GW algorithm performance comparing the output (i.e., the cut-size) of the MaxCut problem using graph partitioning and unpartitioned graph. We show the ratio between the partitioned solution and the unpartitioned solution. Each of the plots shows the performance with different partition sizes $k$, where a larger $k$ represents more partitions used.
• Figure 4: Time it takes to generate batch circuits for QAOA for 5 nodes and varying the number of samples with MC optimization.
• Figure 5: Comparison of state vector simulation with and without Maestro for different shot counts.