| Title |
Reducing Communication Overheads in MD Simulations: A Novel Floating-point Data Compression Approach |
| Authors |
(Seongmin Ki) ; (Sungju Ryu) |
| DOI |
https://doi.org/10.5573/JSTS.2025.25.3.301 |
| Keywords |
Molecular dynamics; LAMMPS; off-chip communication; hardware accelerator; data compression |
| Abstract |
Molecular dynamics (MD) simulations are critical tools for modeling the physical behavior of materials at atomic and molecular scales. However, widely used MD simulators such as LAMMPS [1] suffer from performance when the amount of inter-chip communication increases due to the MPI-based parallelized multi-chip computations at the large simulation data cases. To address this challenge, we propose an efficient data compression method tailored for MD computations. Our technique efficiently compresses the sign and exponent parts of floating-point data (12 bits in the FP64 format) into a single bit, significantly reducing communication bandwidth and energy consumption compared to conventional methods, enhancing the overall energy efficiency of large-scale MD simulations. |