The harpy machine

Abstract

Efficiently storing and retrieving data has always been one of the major problems in computer technology. Until recently there has been more interest for the technology of storing and retrieving data in mass memory than for the problem of accessing data in main memory. Semiconductor technology improvements have lowered main memory cost and increased speed, therefore opening new application areas (e.g. image and speech understanding, graphics) to low cost systems. Some of the tasks in these areas are data intensive and require complex data structures. In these cases, most of the computing power is used to access data rather than to perform operations on them. The performance of such applications can be improved if the memories are tailored to the data structures they contain and to the access operations that can be performed on those data structures. The paper describes an architecture designed to explore the advantages offered by tailored, intelligent memories. The architecture has been implemented and tested on the Harpy speech understanding system. The implementation has been measured and the results are reported in the paper.