GPU-Accelerated Distributed QAOA on Large-scale HPC Ecosystems

Zhihao Xu, Srikar Chundury, Seongmin Kim, Amir Shehata, Xinyi Li, Ang Li, Tengfei Luo, Frank Mueller, In-Saeng SuharXiv 2025 · DOI · arXiv

Introduction

Quantum computers are poised to transform combinatorial optimization, yet we remain in the Noisy Intermediate-Scale Quantum (NISQ) era, where hardware is constrained by qubit counts and error rates. Classical simulators are vital for algorithm validation but face an "exponential wall" in memory and compute requirements as qubit counts grow. This paper presents a GPU-accelerated, distributed approach to the Quantum Approximate Optimization Algorithm (QAOA) tailored for massive HPC systems like the Frontier supercomputer. By integrating advanced problem decomposition with GPU parallelism, the researchers achieved a 10× speedup over CPU-based simulations, enabling optimization of problems with up to 1,000 variables.

Methodology

The scalability bottleneck is addressed through three core pillars:

  1. GPU Acceleration: High-performance simulators like NWQ-Sim and Qiskit Aer are optimized for AMD MI250X GPUs on Frontier. GPUs excel at the matrix-vector multiplications required for quantum state evolution.
  2. Advanced Problem Decomposition: The Impact Factor Directed (IFD) decomposition method identifies high-impact variables through sensitivity analysis, grouping them to preserve critical interactions in sub-problems.
  3. Quantum Framework (QFw): A prototype software stack abstracts HPC complexity, managing circuit distribution across heterogeneous clusters.
The Quantum Framework (QFw) software stack architecture

Performance Evaluation

The approach was evaluated using QASMBench, Max-Cut problems (100–1,000 nodes), synthetic QUBOs, and GHZ state circuits.

Key Results

  • 10× Speedup: GPU-accelerated simulations consistently outperformed multi-core CPU implementations.
  • The GPU "Tipping Point": Computational advantages emerge once problem sizes exceed 16 qubits; below this, data transfer overhead dominates.
  • Strong Scalability: The system scales effectively up to 1,000 cores for large instances.
  • Framework Efficiency: QFw enabled an 11.7× speedup for a 30-qubit GHZ circuit when scaling from 2 to 16 nodes, representing a 91.5% reduction in runtime.
GPU performance advantage for large-scale quantum circuit simulation

Discussion and Future Work

A critical insight is that problem partitioning (IFD) significantly impacts convergence and accuracy compared to random decomposition. However, communication overhead during sub-solution aggregation remains a bottleneck for very large problems, highlighting the importance of high-speed interconnects like HPE Slingshot.

Future work will focus on:

  • Seamlessly integrating DQAOA into the QFw stack.
  • Incorporating noise-aware and density matrix simulations to better mimic real-world quantum hardware.
  • Developing dynamic scheduling for hybrid classical-quantum workloads.
  • Validating the framework on real quantum hardware via QCUP.

Summary updated

BibTeX
@misc{xu2025gpu,
  author = {Zhihao Xu and Srikar Chundury and Seongmin Kim and Amir Shehata and Xinyi Li and Ang Li and Tengfei Luo and Frank Mueller and In-Saeng Suh},
  title = {{GPU-Accelerated Distributed QAOA on Large-scale HPC Ecosystems}},
  year = {2025},
  eprint = {2506.10531},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.10531},
}
Figure