Opportunities and challenges in utilizing application specific integrated circuits in drug design

摘要

Using the MDGRAPE-2 integrated circuits one can reach about two orders of magnitude speedup with respect to the host CPU in typical drug discovery MD simulation provided we do not require periodic boundary conditions. These results predicted by theoretical considerations are demonstrated in practical implementation. The acceleration of methods with either continuum solvent models or with periodic boundary conditions is limited by the time of the CPU-bound subroutines. The maximum speed-up with a single host CPU is about 5 to 25-fold, depending on the method used. While the particle mesh approximation improves the performace of the evaluation of the Ewald sum on CPU-based architectures, the same approximation results in a slow-down on special purpose computing architecture due to the heavy non-ASIC-implementable overhead. Based on these results special purpose computing architectures with a single CPU and up to 12 MDGRAPE-2 chips can provide comparable performance to that of high-end multiprocessor servers with up to 128 CPUs. When periodic boundary conditions or continuum solvent models are necessary, the performance of the same architecture would be more modest. The ideal ratio of CPU and the MDGRAPE-2 chips depend s on the applications.