Development of a Connectionist Network Supercomputer
Krste Asanović, James D. Beck, Jerry Feldman, Nelson H. Morgan, John Wawrzynek · 1993
This paper describes an effort at UC Berkeley and the International Computer Science Institute to develop a super-computer for artificial neural network applications. We describe our applications targets, machine goals, and the system architecture. 1 Introduction The goals of the Connectionist Network Supercomputer-1 (CNS-1) project are to produce super-computer performance and flexible software for connectionist computation at moderate cost. This paper outlines why we have chosen this task and gives the current state of the design. In a university-based research environment, there are two reasons for undertaking an ambitious development project: either the result is needed or one expects the effort to advance science and technology. Both motivations are equally important for CNS-1. Recent work in our lab and elsewhere has shown the practicality of connectionist systems for a range of important problems, but has also revealed needs for computational resources far exceeding those avail...