Force field-based classical molecular dynamics(CMD)is efficient but its potential energy surface(PES)prediction error can be very large.Density functional theory(DFT)-based ab-initio molecular dynamics(AIMD)is accurat...Force field-based classical molecular dynamics(CMD)is efficient but its potential energy surface(PES)prediction error can be very large.Density functional theory(DFT)-based ab-initio molecular dynamics(AIMD)is accurate but computational cost limits its applications to small systems.Here,we propose a molecular dynamics(MD)methodology which can simultaneously achieve both AIMD-level high accuracy and CMD-level high efficiency.The high accuracy is achieved by exploiting deep neural network(DNN)’s arbitrarily-high precision to fit PES.The high efficiency is achieved by deploying multiplication-less DNN on a carefully-optimized special-purpose non von Neumann(NvN)computer to mitigate the performance-limiting data shuttling(i.e.,‘memory wall bottleneck’).By testing on different molecules and bulk systems,we show that the proposed MD methodology is generally-applicable to various MD tasks.The proposed MD methodology has been deployed on an in-house computing server based on reconfigurable field programmable gate array(FPGA),which is freely available at http://nvnmd.picp.vip.展开更多
Molecular dynamics(MD)is an indispensable atomistic-scale computational tool widely-used in various disciplines.In the past decades,nearly all ab initio MD and machine-learning MD have been based on the general-purpos...Molecular dynamics(MD)is an indispensable atomistic-scale computational tool widely-used in various disciplines.In the past decades,nearly all ab initio MD and machine-learning MD have been based on the general-purpose central/graphics processing units(CPU/GPU),which are well-known to suffer from their intrinsic“memory wall”and“power wall”bottlenecks.Consequently,nowadays MD calculations with ab initio accuracy are extremely time-consuming and power-consuming,imposing serious restrictions on the MD simulation size and duration.To solve this problem,here we propose a special-purpose MD processing unit(MDPU),which could reduce MD time and power consumption by about 103 times(109 times)compared to state-of-the-art machine-learningMD(ab initio MD)based on CPU/GPU,while keeping ab initio accuracy.With significantly-enhanced performance,the proposed MDPU may pave a way for the accurate atomistic-scale analysis of large-size and/or longduration problems which were impossible/impractical to compute before.展开更多
基金This work is supported by the National Natural Science Foundation of China(#61804049)the Fundamental Research Funds for the Central Universities of P.R.China+3 种基金Huxiang High Level Talent Gathering Project(#2019RS1023)the Key Research and Development Project of Hunan Province,P.R.China(#2019GK2071)the Technology Innovation and Entrepreneurship Funds of Hunan Province,P.R.China(#2019GK5029)the Fund for Distinguished Young Scholars of Changsha(#kq1905012).
文摘Force field-based classical molecular dynamics(CMD)is efficient but its potential energy surface(PES)prediction error can be very large.Density functional theory(DFT)-based ab-initio molecular dynamics(AIMD)is accurate but computational cost limits its applications to small systems.Here,we propose a molecular dynamics(MD)methodology which can simultaneously achieve both AIMD-level high accuracy and CMD-level high efficiency.The high accuracy is achieved by exploiting deep neural network(DNN)’s arbitrarily-high precision to fit PES.The high efficiency is achieved by deploying multiplication-less DNN on a carefully-optimized special-purpose non von Neumann(NvN)computer to mitigate the performance-limiting data shuttling(i.e.,‘memory wall bottleneck’).By testing on different molecules and bulk systems,we show that the proposed MD methodology is generally-applicable to various MD tasks.The proposed MD methodology has been deployed on an in-house computing server based on reconfigurable field programmable gate array(FPGA),which is freely available at http://nvnmd.picp.vip.
基金supported by the National Natural Science Foundation of China(62474058 and 61804049)the Yuelushan Center for Industrial Innovation(2023YCII0104)+2 种基金the Huxiang High Level Talent Gathering Project(2019RS1023)the Technology Innovation and Entrepreneurship Funds of Hunan Province,P.R.China(2019GK5029)the Fund for Distinguished Young Scholars of Changsha(kq1905012).
文摘Molecular dynamics(MD)is an indispensable atomistic-scale computational tool widely-used in various disciplines.In the past decades,nearly all ab initio MD and machine-learning MD have been based on the general-purpose central/graphics processing units(CPU/GPU),which are well-known to suffer from their intrinsic“memory wall”and“power wall”bottlenecks.Consequently,nowadays MD calculations with ab initio accuracy are extremely time-consuming and power-consuming,imposing serious restrictions on the MD simulation size and duration.To solve this problem,here we propose a special-purpose MD processing unit(MDPU),which could reduce MD time and power consumption by about 103 times(109 times)compared to state-of-the-art machine-learningMD(ab initio MD)based on CPU/GPU,while keeping ab initio accuracy.With significantly-enhanced performance,the proposed MDPU may pave a way for the accurate atomistic-scale analysis of large-size and/or longduration problems which were impossible/impractical to compute before.
