Massively parallel computation of atmospheric neutrino oscillations on CUDA-enabled accelerators

The computation of neutrino flavor transition amplitudes through inhomogeneous matter is a time-consuming step and thus could benefit from optimization and parallelization. Next to reliable parameter estimation of intrinsic physical quantities such as neutrino masses and mixing angles, these transition amplitudes are important in hypothesis testing of potential extensions of the standard model of elementary particle physics, such as additional neutrino flavors. Hence, fast yet precise implementations are of high importance to research. In the recent past, massively parallel accelerators such as CUDA-enabled GPUs featuring thousands of compute units have been widely adopted due to their superior memory bandwidth, vast compute capability, and highly competitive compute-to-energy ratio in comparison to traditional multi-core architectures with a few tens of monolithic cores. In this paper, we introduce two scalable multi-GPU extensions of common neutrino oscillation frameworks –namely Prob3++ and SQuIDS –allowing for the acceleration of oscillation dynamics computation by one to three orders-of-magnitude while preserving numerical accuracy. Our software is licensed under LGPLv3 and can be accessed at https://github.com/fkallen/CUDAProb3 and https://github.com/fkallen/CUDAnuSQuIDS.

通过非均匀介质（inhomogeneous matter）计算中微子味跃迁振幅（neutrino flavor transition amplitudes）是一项耗时的工作，因此可通过优化与并行化技术提升计算效率。除了对中微子质量、混合角等内禀物理量开展可靠的参数估计之外，这类跃迁振幅在检验粒子物理标准模型（standard model of elementary particle physics）潜在扩展理论（如引入额外中微子味）的假说时也具备重要价值。因此，兼具快速性与精确性的实现方案对相关研究具有极高的重要性。近年来，搭载数千个计算单元的支持CUDA的图形处理器（GPU）这类大规模并行加速器得到了广泛应用——相较于仅配备数十个整体式核心的传统多核架构，此类加速器拥有更优异的内存带宽、更强的计算能力，以及更具竞争力的计算能效比。本文针对两类主流的中微子振荡框架——即Prob3++与SQuIDS——提出了可扩展的多GPU扩展方案，可将振荡动力学计算的速度提升1至3个数量级，同时保留数值精度。本软件采用LGPLv3许可证开源，可通过以下链接获取：https://github.com/fkallen/CUDAProb3 与 https://github.com/fkallen/CUDAnuSQuIDS。