Research And Design Of Molecular Dynamics Computing System Based On FPGA

Molecular Dynamics(MD)simulation has been one of the important methods used by researchers in the past decades to study macroscopic properties of materials from the microscopic perspective.This method can study the physical and chemical properties of new materials and drugs from the microscopic perspective.However,due to the computational complexity of their algorithms,a simple simulation of a biological entity on a conventional processor usually takes months.FPGA stands out among many acceleration platforms due to its low power consumption and high energy efficiency.However,the traditional RTL level development of FPGA has great programming difficulty and engineering design cost,which limits the application of FPGA in MD simulation acceleration.Aiming at the above problems,this paper proposed a molecular dynamics acceleration method based on FPGA.The design idea of HLS+HDL effectively solved the problems of too long MD simulation time and the low efficiency of HLS in the implementation of dynamic data flow.Finally,based on this design idea,the acceleration design of two potential energy algorithms was realized.The main work and innovation of this paper are as follows:Firstly,the dynamic data path existing in MD is analyzed,and the reason why HLS cannot be efficiently described is obtained.Then,a design method of HLS+HDL is proposed,which combines the high efficiency of HLS development with the refinement of HDL development to greatly reduce the design time and ensure the performance of the accelerator.Secondly,aiming at the two-body potential energy Lj,liquid argon is selected as the simulation object of MD.Based on the open source simulation software LAMMPS,a FPGAbased accelerator is designed to realize dynamic data flow in MD by using HLS+HDL,and optimization strategies such as efficient data cache design and calculation pipeline design are adopted.It not only greatly reduces the design time and engineering cost of the whole system,but also achieves good performance.The experimental results show that for 20 K liquid argon simulation object,the reconfigurable accelerator based on Xilinx U50 can achieve 200.73ns/day of simulation performance.Compared with Intel Xeon E5-2620,the energy efficiency of a single node is 159.28 times,and the simulation performance is 25.8 times.Compared with the SW26010 single core group with 4 main cores and 256 slave cores carried by the supercomputer Sunway Taihu Light,the energy efficiency of a single node is 17.07 times higher and the simulation performance is 1.61 times higher.Thirdly,for the three-body potential energy Tersoff,C is selected as the MD simulation object,and a FPGA-based accelerator is designed based on LAMMPs.Several cache strategies are carefully designed and analyzed.Finally,according to the calculation characteristics of Tersoff,a full pipeline design is carried out to achieve the maximum performance.Experimental results show that for 512 K carbon atoms,the reconfigurable accelerator based on Xilinx U50 can achieve a simulation performance of 1.87ns/day.Compared with Intel Xeon E5-2620,the energy efficiency of a single node is 91.8 times,and the simulation performance is 26.2times.Compared with SW26010 single core group,the energy efficiency of single node is 9.9times and simulation performance is 1.19 times.