[Paper Review] Hybrid quantum programming with PennyLane Lightning on HPC platforms

Ali Asadi, Amintor Dusko+6 more|arXiv (Cornell University)|Mar 4, 2024

Quantum Computing Algorithms and Architecture4 citations

TL;DR

This paper introduces PennyLane Lightning, a high-performance suite of C++-backed state-vector simulators optimized for CPU, GPU, and HPC environments. It enables efficient hybrid quantum-classical workloads—such as VQE, QAOA, and gradient-based optimization—through SIMD acceleration, multi-threading, batched execution, and distributed MPI-based simulations across multiple nodes, achieving up to 41-qubit circuit simulations with strong performance gains over existing simulators, especially on GPU and multi-node HPC platforms.

ABSTRACT

We introduce PennyLane's Lightning suite, a collection of high-performance state-vector simulators targeting CPU, GPU, and HPC-native architectures and workloads. Quantum applications such as QAOA, VQE, and synthetic workloads are implemented to demonstrate the supported classical computing architectures and showcase the scale of problems that can be simulated using our tooling. We benchmark the performance of Lightning with backends supporting CPUs, as well as NVidia and AMD GPUs, and compare the results to other commonly used high-performance simulator packages, demonstrating where Lightning's implementations give performance leads. We show improved CPU performance by employing explicit SIMD intrinsics and multi-threading, batched task-based execution across multiple GPUs, and distributed forward and gradient-based quantum circuit executions across multiple nodes. Our data shows we can comfortably simulate a variety of circuits, giving examples with up to 30 qubits on a single device or node, and up to 41 qubits using multiple nodes.

Motivation & Objective

• To develop a high-performance, portable quantum simulator suite that supports diverse classical hardware backends including CPUs, NVIDIA and AMD GPUs, and HPC clusters.
• To enable efficient execution of hybrid quantum-classical workloads such as variational quantum algorithms (VQE, QAOA) and gradient-based optimization using automatic differentiation.
• To support scalable, distributed quantum circuit simulation across multiple nodes using MPI, enabling simulation of large circuits beyond single-device limits.
• To provide performant, natively integrated quantum gradient computation (via adjoint method) across all backends, including distributed GPU environments.
• To offer researchers accurate runtime estimation guidelines for resource allocation in HPC environments, facilitating efficient workload planning for large-scale quantum simulations.

Proposed method

• Design and implement a modular C++20-based simulator stack with optimized gate kernels using SIMD intrinsics (AVX-512, AVX2) for improved single- and multi-threaded performance on CPUs.
• Implement device backends for lightning.qubit (CPU), lightning.gpu (NVIDIA CUDA), and lightning.kokkos (AMD ROCm) with unified interface via PennyLane’s device-agnostic framework.
• Enable batched, task-based execution of independent circuits across multiple GPUs using CUDA streams and multi-GPU task scheduling.
• Implement distributed state-vector simulation using MPI for multi-node execution, supporting both forward and gradient-based circuit evaluations across clusters.
• Integrate native support for adjoint-based Jacobian computation across all backends, enabling efficient reverse-mode automatic differentiation in JAX, PyTorch, and TensorFlow workflows.
• Leverage modern C++20 features and abstraction layers to ensure portability and performance across diverse HPC platforms, including Perlmutter, LUMI, and AWS EC2.

Experimental results

Research questions

• RQ1 How does PennyLane Lightning’s performance compare to other high-performance simulators in single-node, multi-GPU, and distributed multi-node HPC environments?
• RQ2 To what extent can SIMD intrinsics and multi-threading improve CPU-based quantum circuit simulation performance for variational algorithms?
• RQ3 What is the scalability and efficiency of distributed quantum circuit simulation using MPI across multiple nodes and GPUs?
• RQ4 Can the adjoint method for gradient computation be efficiently and natively supported across CPU, GPU, and distributed HPC backends in a unified framework?
• RQ5 How can researchers estimate resource requirements and runtime for large-scale quantum simulations using this toolset in HPC environments?

Key findings

• Lightning achieves up to 30% stronger weak scaling efficiency on 512 GPUs for gradient-based workloads (e.g., circuit gradients) compared to other simulators, with 8% and 30% efficiency for sampling and gradient workloads, respectively.
• On 30-qubit circuits, Lightning’s CPU backend with AVX-512 shows significant performance improvements over baseline implementations, particularly in gate application time, outperforming other packages for specific gate types.
• The GPU backend (lightning.gpu) demonstrates a performance advantage over CPU for circuits beyond 20 qubits, with increasing benefit as circuit depth and qubit count grow.
• Lightning.gpu is the only distributed state-vector simulator that natively supports the adjoint method for gradient computation across multiple nodes, enabling efficient reverse-mode differentiation in large-scale simulations.
• The framework enables simulation of circuits up to 41 qubits using distributed multi-node execution, demonstrating feasibility for large-scale quantum algorithm prototyping.
• Runtime estimation guidelines are provided for various workloads, helping researchers plan resource allocation on HPC systems such as Perlmutter and LUMI, with validated benchmarks using real HPC hardware.