Design And Research Of CPU-FPGA Based High-speed Molecular Dynamics Heterogeneous Computing System

The Molecular Dynamics(MD)simulation method starts from the microscopic scale and uses the microscopic states generated by the trajectories of atomic motion,averaged over time,to obtain the kinetic properties of the studied system.Molecular dynamics has become an important tool as a powerful research tool to assist and even guide experiments,and is widely used in the fields of materials,medicine,chemistry,and geology.However,the complex computational approach of MD method makes it take months to simulate a simple viral structure.FPGA,as a non-Von Neumann computing architecture,can provide higher parallel computing power compared with the computing architecture of traditional processors,which is ideal as a platform for MD simulation computation.However,the complex mathematical equations in MD methods are extremely difficult to implement in FPGAs,which also hinders FPGAs from becoming the mainstream MD computing platform.For the above problems,this paper proposes a CPU-FPGA based high-speed molecular dynamics heterogeneous computing system that uses deep neural networks to fit the force field model in MD,which effectively overcomes the problem that the classical force field model is difficult to deploy in FPGAs,while the heterogeneous model proposed in this paper improves the speed of calculating the forces by 11 times compared to the computational speed of traditional CPUs.The main work and innovations of this paper are as follows.First,the structure of the Neural Network Force Field(NNFF)model is analyzed,and the implementation form of the NNFF model in FPGA hardware is deeply optimized.The model compression method is used to quantize the network parameters in floating-point form into fixed-point form,which significantly reduces the usage of FPGA logic resources without reducing the model accuracy.An activation function with a shape similar to the tanh function is implemented by the quadratic term approximation method,which effectively overcomes the problem that the exponential term in the traditional activation function is difficult to implement in FPGAs and enables the model to be easily deployed to FPGAs.Second,the quantized neural network is used to fit the force field of crystalline silicon(Si64)to obtain a NNFF with a network size of 120×30×30×3 and 9-bit fixedpoint network parameters,which has a root-mean-square error of 0.0841 eV/(?).The trained quantized NNFF model is written in RTL code and designed to work in a fully pipelined manner,using Vivado software for functional simulation,and the root mean square error of the force calculation accuracy is 0.0854 eV/(?).Finally,the whole CPU-FPGA high-speed molecular dynamics heterogeneous computing system was designed by analyzing the execution logic of the MD method.The Linux-based PCIe driver and XDMA IP core implement the heterogeneous communication between CPU and FPGA;the feature construction module and feature encoding module of atoms are implemented in the CPU;and the NNFF module for calculating atomic forces is implemented in the FPGA.Finally,through practical tests,the heterogeneous system makes the computation of stresses nearly 11 times faster compared to the pure CPU platform.