利用GPU进行电磁暂态仿真的时间特性与关键参数分析

Analysis on Temporal Characteristics and Key Parameters of Electromagnetic Transient Simulation Accelerated by GPU

中文关键词:电磁暂态  仿真  矩阵-向量相乘

keywords:electromagnetic transient  simulation  matrix-vector multiplication

中文摘要:现有关于图形处理器（graphics processing unit，GPU）加速电磁暂态（electromagnetic transient，EMT）仿真的研究，缺乏对时间特性的系统量化分析。针对这一问题，建立GPU执行矩阵-向量乘法的计算时间模型，将总计算时间分解为基本开销项、计算规模与带宽主导项2部分。基于3种不同架构GPU的时间相关数据完成模型参数量化。实验结果显示：GPU带宽利用率可达0.83~0.90，而算力利用率显著偏低，计算时间完全受显存带宽限制；基本开销约为0.3~0.5 ms，且随带宽增加略有增大，这一开销导致 GPU 难以满足微秒级步长的 EMT 实时仿真需求；显存容量决定可支持的最大矩阵规模，接近容量上限时会显著影响计算性能。所提模型与量化结果可用于预测 GPU 在 EMT仿真中的性能边界，为相关计算平台设计提供参考。

Abstract:Existing studies on GPU-accelerated electro-magnetic transient （EMT） simulation lack systematic quantitative analysis of temporal characteristics.To solve this problem，this paper establishes a computation time model for matrix-vector multiplication on GPU，which splits the total computation time into two components:base overhead term and dominant term determined by computation scale and memory bandwidth.Model parameters are quantified based on time-dependent data of three GPUs with different architectures.The results indicate that GPU memory bandwidth utilization reaches 0.83-0.90，while compute resource utilization remains considerably low； the overall computation time is fully limited by video memory bandwidth.The base overhead is approximately 0.3-0.5 ms and rises slightly with higher bandwidth.Such overhead makes GPUs unable to meet the requirement of real-time EMT simulation with microsecond-level time steps.In addition，video memory capacity determines the maximum supported matrix size，and computation performance degrades sharply when the memory usage approaches the capacity limit.The proposed model and quantitative results can be used to predict the performance boundary of GPUs in EMT simulation and provide references for the design of relevant computing platforms.