Computation with Serial Memory Systems.
Abstract
This paper examines the question of whether it is possible to substitute serial memory for conventional random access memory (RAMs) in commonly encountered numerical computations. There is much interest in tasks such as solving large systems of linear equations, large linear programming problems, and other calculations applied to large amounts of data concurrently with development of techniques for constructing large serial memories which are intrinsically less expensive than comparable RAMs. The answer to the question is 'yes' for those computations discussed in this paper if the serial memory is equipped with compare/exchange circuits so that it will sort its contents. Generally when using self-sorting memory, labels or keys are appended to the data to be stored; conventional machine addresses are not to be used. Data entries thus specified pass to simple arithmetic processors for the requisite numerical operations, after which the data may be returned to the self-sorting memory for retention until needed again. A property common to all computations discussed in this paper is the feasibility of computing the requisite memory output sequence as data is first passed to the memory, as well as in the course of computation. While there is clearly no analytic basis for asserting that the class of all common numerical computations shares this property, many computations not discussed in this paper are clearly within this class.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 01, 1979
- Accession Number
- ADA068137
Entities
People
- Martin Rem
- Philip N. Armstrong