Toward an FPGA-based dedicated computer for molecular dynamics simulations.

Area of Science:

• Computational Physics
• Materials Science
• High-Performance Computing

Background:

• Molecular dynamics (MD) simulations are crucial for understanding material properties and processes.
• Current computational resources face limitations in simulating large-scale or long-timescale phenomena.
• Field-Programmable Gate Arrays (FPGAs) offer potential for specialized, high-speed scientific computing.

Purpose of the Study:

• To present initial steps for implementing molecular dynamics (MD) on a Field-Programmable Gate Array (FPGA) cluster.
• To evaluate the simulation speed and scalability of an FPGA-based MD system.
• To demonstrate the feasibility of FPGA clusters for multi-microsecond simulations, including complex phenomena like nucleation.

Main Methods:

• Development of a highly parallelized and pipelined MD algorithm tailored for FPGA implementation.
• Utilizing a cluster of Artix 7 XC7A200T FPGAs interconnected via fast optical links in a 3D torus topology.
• Focus on optimizing high-speed networking and implementing essential MD functionalities (Lennard-Jones interactions, thermostat).

Main Results:

• Achieved a simulation speed of a few microseconds per day, demonstrating significant progress in FPGA-based MD.
• Confirmed the scalability of the FPGA cluster, allowing for larger systems without compromising simulation speed.
• Successfully simulated the nucleation of a super-cooled Lennard-Jones liquid, validating the system's capability for challenging problems.

Conclusions:

• The presented FPGA cluster architecture is feasible for accelerating multi-microsecond molecular dynamics simulations.
• The system demonstrates excellent scalability and potential for future, more comprehensive MD implementations.
• The developed MD implementation will be released as an open-source project to foster further research and development.