Design Considerations for Hardware Based Acceleration of Molecular Dynamics

As demand for long and accurate molecular simulations increases so too does the computation demand. Beyond using new, enterprise scale processor developments - such as the ARM neoverse chips � or performing simulations leveraging Graphics Processing Unit compute, there exists a potentially faster and more power efficient option in the form of custom hardware. Using hardware description languages it is possible to transform existing algorithms into custom, high performance hardware layouts. This can lead to faster and more efficient simulations but compromises on the required development time and flexibility. In order to take the greatest advantage of the potential performance gains, the focus should be on transforming the most computationally expensive parts of the algorithms. When performing molecular dynamics simulations in a polar solvent like water, non-bonded electrostatic calculations dominate each simulation step, as the interactions between the solvent and the molecular structure are calculated. However, simply developing a non-bonded electrostatics co-processor may not be enough, as transferring data between the host program and the Field Gate Programable Arrays itself incurs a significant time delay. For any changes to be made, competitive to existing calculation solutions, the number of data transfers must be reduced. This could be achieved by simulating multiple time-steps between memory transfers, which may impact accuracy, or performing more calculations in the custom hardware.