[Paper Review] Simulating Quantum Computers Using OpenCL

Adam Kelly|arXiv (Cornell University)|May 2, 2018

Quantum Computing Algorithms and ArchitectureComputer Science7 references20 citations

TL;DR

This paper presents QCGPU, an OpenCL-based quantum circuit simulator that leverages GPU acceleration to significantly outperform existing state vector simulators. By implementing optimized kernel launches and parallel gate application via OpenCL, the simulator achieves over 150x speedup over Qiskit and 8x over ProjectQ at 24 qubits, demonstrating the effectiveness of portable hardware acceleration for quantum algorithm development.

ABSTRACT

Quantum computing is an emerging technology, promising a paradigm shift in computing, and allowing for speedups in many different problems. However, quantum devices are still in their early stages, most with only a small number qubits. This places a reliance on simulation to develop quantum algorithms and to verify these devices. While there exists many algorithms for the simulation of quantum circuits, there is (at the time of writing) no tools which use OpenCL to parallelize this simulation, thereby taking advantage of devices such as GPUs while still remaining portable. In this paper, such a tool is described, including optimizations in areas such as gate application. This leads to a new approach that outperforms other popular state vector based simulators. An implementation of the proposed simulator is available at https://qcgpu.github.io.

Motivation & Objective

• To address the lack of portable, hardware-accelerated quantum circuit simulators that leverage GPUs.
• To enable efficient simulation of quantum algorithms on classical hardware, especially as quantum devices remain limited in qubit count.
• To develop a portable, open-source simulator using OpenCL for cross-platform GPU acceleration.
• To outperform existing state vector simulators through optimized kernel design and parallel execution.
• To lower the barrier to entry for researchers by providing a freely available, high-performance simulation tool.

Proposed method

• Implementing a quantum circuit simulator using OpenCL to target heterogeneous hardware including GPUs.
• Designing compute kernels in OpenCL C to perform parallel state vector updates across qubit gates.
• Optimizing kernel launch configurations with work groups and work items to maximize GPU occupancy and memory coalescing.
• Using a host-driven pipeline to manage memory transfers between CPU and GPU and to schedule kernel execution.
• Applying algorithmic optimizations such as efficient gate application and memory access patterns to reduce latency.
• Benchmarking against Qiskit and ProjectQ using standard quantum Fourier transform circuits across 1–24 qubits.

Experimental results

Research questions

• RQ1 Can OpenCL-based hardware acceleration significantly improve the performance of quantum circuit simulation compared to existing CPU-only simulators?
• RQ2 How does the performance of an OpenCL-based simulator scale with increasing qubit count, especially beyond 20 qubits?
• RQ3 To what extent does the choice of kernel launch configuration and memory access pattern affect simulation efficiency on GPU architectures?
• RQ4 Can a portable, cross-platform solution using OpenCL achieve performance comparable to or better than vendor-specific GPU frameworks?
• RQ5 What are the practical limits of single-device GPU simulation for quantum circuits, and how does it compare to distributed or tensor network methods?

Key findings

• QCGPU achieved an average speedup of over 150 times compared to Qiskit at 24 qubits for quantum Fourier transform circuits.
• The simulator outperformed ProjectQ by a factor of 8x on average at the 24-qubit level.
• Statistical analysis using Welch’s t-test confirmed the performance difference was significant (p < 0.001) against both Qiskit and ProjectQ.
• The simulator successfully simulated up to 28 qubits on a single GPU, demonstrating scalability within hardware limits.
• The use of OpenCL enabled portable, high-performance simulation across diverse GPU architectures without vendor-specific code.
• The benchmarking results showed consistent performance gains across all tested qubit counts, with the largest relative improvement observed at higher qubit counts.