Error evaluation in the design of a special‐purpose processor that calculates nonbonded forces in molecular dynamics simulations

Abstract

Listen

Translate article

AbstractSpecial‐purpose parallel machines that are plugged into a workstation to accelerate molecular dynamics (MD) simulations are attracting a considerable amount of interest. These machines comprise scalable homogeneous multiprocessors for calculating nonbonded forces (Coulombic and van der Waals forces), which consume more than 99% of the central processing unit (CPU) time in standard MD simulations. Each processor element in the machine has a pipeline architecture to calculate the total nonbonded force exerted on a particle by all of the other particles using information regarding the coordinates, the electric charge, and the species of each particle broadcast by the host computer. The processor then sends the calculated force back to the host computer. This article addresses the precision of the calculated nonbonded forces in the design of a processor LSI with minimal complexity. The precision of the arithmetic inside the processor that is required to calculate forces for MD simulations using Verlet's procedure was critically evaluated. Forward and backward error analysis, coupled with numerical MD experiments on one‐dimensional systems, was performed, and the following results were obtained: (1) Each element of the position vector which the processor receives from the host computer should have a precision of at least 25 bits; and (2) the pairwise forces should be calculated using floating point numbers with at least 29 bits of mantissa in the processor. Calculation of a pairwise force, which involves second‐order polynomial interpolation using a table‐driven algorithm, requires a key which contains a duplicate of at least 11 most significant bits of mantissa of the squared pairwise distance. The final result was that (3) the total force that acts on a particle, which is obtained by summing the forces exerted by all of the other particles, should be calculated using an accumulator that has a mantissa of at least 48 bits. © 1995 by John Wiley & Sons, Inc.