FPGA-Accelerated Tersoff Multi-body Potential for Molecular Dynamics Simulations

• Published in: Applied Reconfigurable Computing. Architectures, Tools, and Applications Vol. 13569; pp. 17 - 31
• Main Authors: Yuan, Ming, Liu, Qiang, Deng, Quan, Xiang, Shengye, Gan, Lin, Yang, Jinzhe, Duan, Xiaohui, Fu, Haohuan, Yang, Guangwen
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer 2022; Springer Nature Switzerland
• Series: Lecture Notes in Computer Science
• Subjects: Accelerator; FPGA; Molecular dynamics simulations; Pipeline
• ISBN: 9783031199820; 3031199820
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-031-19983-7_2

Molecular Dynamics simulation (MD) models the interactions of thousands to millions of particles through the iterative application of fundamental physics, and MD is one of the core methods in High-Performance Computing (HPC). However, the inherent weak scalability problem of force interactions renders MD simulation quite computationally intensive and challenging to scale. To this end, specialized FPGA-based accelerators have been proposed to solve this problem. In this work, we focus on many-body potentials on a single FPGA. Firstly, we proposed an efficient data transfer strategy to eliminate the latency between on-chip and off-chip memory. Then, the fixed-point description of data type is developed for computation to increase the utilization of on-chip resources. At last, a custom pipelined strategy is presented for Tersoff to get a better simulation performance. Compared with a floating-point implementation based on NVIDIA 28080ti GPUs, our design based on Xilinx U200 FPGA is 1.2 times better.