中文摘要：

作为高性能科学计算的典型应用，利用GPU并行加速分子动力学模拟是2007年以来计算化学领域高性能计算的热点。本文概述了支持GPU加速的不同MD软件的特点和其研究进展，重点分析了Amber、GROMACS、ACEMD三个代表性软件的单GPU卡和多GPU卡计算性能，结果表明在配置相同数目GPU卡的情况下，单节点比多节点在计算性能上较有优势，桌面工作站配多块GPU卡是性价比相对较好的MD模拟计算模式。本文还考察了单精度和双精度GPU加速MD的模拟计算结果的准确性，与CPU的计算结果进行了比较，结果表明，GPU的计算结果总体而言是可信的。最后，本文对GPU并行加速MD模拟的研究现状进行总结并对未来发展做了展望。

英文摘要：

As a typical application of high performance computing, development of parallel code for molecular dynamics （MD） simulations using graphical processing units （GPU） is a hot topic in computational chemistry since 2007. Rapid progress was achieved from the earlier attempts on creating GPU-enabled module for non-bonded interactions such as van der Waals potential or electrostatic interactions, the bottle-neck step in MD simulations, to the MD code fully running on GPU. This paper reviews the GPU-enabled state-of-the-art features of MD implementations in widely used molecular dynamics packages or platforms. The computing performances of three representative MD software, Amber, GROMACS and ACEMD, running on single and multiple GPU cards were benchmarked using de facto standard simulation of DHFR. When the number of GPU cards is fixed, the computing performance of multi-GPU on multi-node is no better than multi-GPU on a single node due to the time-consuming communication between multi-node. Further improving the computing performance of multi-GPU on multi-node is very important for further development of GPU accelerated MD software for simulation of larger molecular model and longer time scale. The benchmark support our point of view that it is a good and cost-effective choice to run GPU accelerated MD simulations on desktop workstations. This paper also addresses another important issue in GPU parallel computing by comparing the GPU-enabled MD simulations using floating point calculations of single-precision, double-precision as well as fixed-precision with CPU using double-precision, which suggest that GPU-accelerated MD simulations using different precision of floating point calculations can be acceptable and depend on applications. Lastly, perspectives of code development for GPU-enabled molecular dynamics are briefly discussed.