Special Purpose Processors for Computing Materials Properties

Abstract

ABSTRACTThe need for computational power in the modeling of physical systems is rapidly increasing. Realistic simulations of materials often require complex interactions and large numbers of particles. For most scientists, full-time access to supercomputers is not possible, and even this might not be sufficient to solve their problems. As most of the calculations involved are straightforward and repetitive in nature, a possible solution is to design and build a processor for a specific application with a low cost/performance ratio. This approach is to be contrasted with the use of a general purpose computer, which is designed to treat a large class of problems and includes many expensive features (e.g. software) that are not utilized in the simulations. The architecture of a special purpose computer can be tailored to the problem; e.g., parallel and pipelined operations can be incorporated to obtain efficient computational throughput, and memory organization and instruction sets can be optimized for this purpose. A few of such algorithm oriented processors have been built in the last decade and have been utilized for certain specific jobs; for example: molecular dynamics simulations of systems of Lennard-Jones particles and Monte Carlo calculations of Ising models. An overview of some existing algorithm oriented processors and expectations for the future will be presented.