在相对论耦合簇理论的张量收缩中利用浮点仿真技术
Exploiting floating point emulation in tensor contractions in relativistic coupled cluster theory
AI总结:
本研究提出利用Ozaki方案I的浮点仿真技术优化TTGT张量收缩模式,可在NVIDIA Blackwell B200等加速器上加速相对论耦合簇理论的关键张量收缩,且支持灵活调整计算精度。
摘要:
在本手稿中,我们概述了一种计算策略,该策略利用针对浮点仿真的Ozaki方案I的供应商优化实现,将其作为张量收缩的Transpose-Transpose-GEMM-Transpose(TTGT)模式实现的一部分,并评估其在高端加速器(如NVIDIA Blackwell B200系统)中加速相对论耦合簇理论中两个最耗时的张量收缩的潜力,该系统中双精度与低精度数据类型的计算性能存在显著差异。我们的结果表明,这种基于仿真的TTGT策略相比无仿真运行可提供显著的加速效果,且可与使用原生双精度作为计算数据类型的最先进供应商优化张量收缩库相媲美。更重要的是,我们的方案为用户提供了一种途径,可轻松调整张量收缩步骤中所需的精度,包括应用中仅需单精度与双精度之间精度的情况。
英文摘要:
In this manuscript we outline a computational strategy to leverage a vendor-optimized implementation of the Ozaki scheme I for floating point emulation as part of an implementation of the Transpose-Transpose-GEMM-Transpose (TTGT) pattern for tensor contractions, and assess its potential to accelerate two of the most expensive tensor contractions appearing in relativistic coupled cluster theory in high-end accelerators such as the NVIDIA Blackwell B200 system in which there is a strong differential in the performance of computations for double and lower precision datatypes. Our results indicate that this emulation-based TTGT strategy can provide significant speedups compared to runs without emulation, and can be competitive with state-of-the-art, vendor-optimized tensor contraction libraries using native double precision as compute datatype. More importantly, our scheme provides a path towards easily allowing users to tune the desired accuracy in the tensor contraction steps, including cases where precision between single and double is sufficient for the application.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。