System and method for achieving improved accuracy from efficient computer architectures

Abstract

This invention provides a system and method that can employ a low-instruction-per-second (lower-power), highly parallel processor architecture to perform the low-precision computations. These are aggregated at high-precision by an aggregator. Either a high-precision processor arrangement, or a low-precision processor arrangement, employing soft-ware-based high-precision program instructions performs the less-frequent, generally slower high-precision computations of the aggregated, more-frequent low-precision computations. One final aggregator totals all low-precision computations and another high-precision aggregator totals all high-precision computations. An equal number of low precision computations are used to generate the error value that is subtracted from the low-precision average. A plurality of lower-power processors can be arrayed to provide the low-precision computation function. Alternatively a plurality of SIMD can be used to alternately conduct low-precision computations for a predetermined number of operations and high-precision operations on a fewer number of operations. In an embodiment, aggregation can include summing values within predetermined ranges of orders of magnitude, via an adding tree arrangement, so that significant digits therebetween are preserved.