Improved Interpolation and System Integration for FPGA-Based Molecular Dynamics Simulations

Abstract

FPGA-based acceleration of molecular dynamics (MD) has been the subject of several recent studies. Here we describe a new non-bonded force computation pipeline implemented on a 2004era COTS FPGA board and its integration into the ProtoMol MD code. There are several innovations: a novel interpolation strategy; the introduction of a "semi-floating point" format; and various issues related to system integration. As a result, we are able to model far more particle types, without relying on complex buffering, and obtain higher accuracy than previously. A two pipeline accelerator has been implemented on a Xilinx VirtexII Pro VP70, integrated into ProtoMol, and tested with an enzyme inhibitor model having 8000 particles and 26 particle types. Despite performing all O(n) work on the host PC, as well as the data conversion and communication overhead, this implementation yields a 5.5x speed-up over a 2.8GHz PC, and with accuracy comparable to the serial code.

FPGA-Accelerated Molecular Dynamics Simulations System

“…Computing the short-range force is with lookup table [12] and interpolation. We use (7) instead of (2) to avoid square root operation.…”

Integrating FPGA Acceleration into the Protomol Molecular Dynamics Code: Preliminary Report

“…The other advantages include the integration of higher order interpolation and the use of "semi floating point." These combined optimizations result in substantial savings in area over double precision floating point, but with little if any sacrifice in simulation accuracy [3].…”

Towards production FPGA-accelerated molecular dynamics: Progress and challenges

“…Our MD force pipelines have been extended to now support the short-range part of the Particle Mesh Ewald method of computing the electrostatic potential (in addition to the Multigrid method previously implemented [13]). This has necessitated a substantial redesign: we now use table lookup with interpolation [7] rather than direct computation [18]. In addition, in order to improve energy fluctuation we have added a switching function to the van der Waals calculation.…”

“…Shaw [4]). Although there have been many FPGA implementations (e.g., [5], [6], [7], [8], [9], [10], [11]), the heavy use of floating point in MD appeared to make these less competitive. In recent work [12], [13], however, we have shown that when FPGAs perform particle filtering (on-the-fly neighbor list generation), much of the floating point computation can be eliminated.…”