Supercomputers = Colossal Computations + Enormous Expectations + Renowned Risk

Publisher Website
Google Scholar
Abstract
A natural focus for any new discussion of supercomputer development - and the central concern of this article - is the superprocessor architecture function. Are supercomputer developers following the same old refrain or creating a new tune? The Control Data Corporation (CDC) Cyber 205 project serves as a basis for this look at current and next-generation requirements. With the Cyber 205, the thinking moved in a full circle. First, overspecialization was evident in the development of the predecessor Star-100 by the Lawrence Livermore Laboratory; then, after some reflection, in part on the exorbitant cost of software development, the scope of the Star-100 was redefined. The new approach incorporated a better balance between vector and scalar performance in a Star-100 type architecture, while relinquishing chores such as communications and unit record management to front-end processors. This revised front-end concept then demanded the development of new interconnect schemes between alien processors to support the data rates needed by a supercomputer. The result of this process was the creation of a new CDC product, called the Loosely Coupled Network (LCN), which is now employed in other environments in addition to supporting the Cyber 205. The success of this scheme has established a clear direction for the next generation of supercomputer systems, in which a hierarchy of processing power is linked via the LCN with one or more Cyber 205 successors.

This publication has 6 references indexed in Scilit: