Getting a group of people to work together efficiently is ademanding task that often requires a skilled manager. Computerdesigners and programmers are now looking for the electronic equivalentof such a manager to cope with “multiprocessor” computers.Each of the many processors built into this type of computer canindependently retrieve data from memory locations, do arithmeticoperations like addition or perform other simple operations.
Yet thecomputer must quickly come up with a “consensus”–a single,correct answer. Last week, the Department of Energy (DOE) and theNational Science Foundation awarded grants totaling $9 million to theUniversity of Illinois at Urbana-Champaign to pursue this problem. David Kuck, director of the university’s new Center forSuper-computing Research, and his colleagues plan to build anexperimental supercomputer, called the “Cedar” system, to testtheir ideas. The problem is to connect the computer hardware andsoftware in such a way that all the multiprocessing occurs in anorganized, productive and fast way, says Kuck. This project is unusualbecause, to achieve high speeds, it relies on clever ways of linking theprocessors and then writing computer programs that take advantage ofthese arrangements rather than on advances in computer chip technology.
DOE also helps fund two other large supercomputing researchprograms. At the California Institute of Technology in Pasadena,Charles L. Seitz, Geoffrey Fox and their colleagues aim to build acomputer that fits on a desktop but still has 50 times the power of aCray-1 supercomputer at a fraction of the Cray’s current price. Theresearchers already have one experimental machine, completed lastOctober and dubbed “The Cosmic Cube,” in which 64 identicalmicroprocessors are connected in a communications network based on asix-dimensional cube. At New York University’s Courant Institute of MathematicalSciences in New York City, researchers are working on the”ULTRAcomputer” project. In this “dataflow”computer, hundreds of thousands of processors share one massive memory,communicating with it through a switching network that keeps theprocessors synchronized. Instead of searching the memory for the dataneeded to do a certain computation, each processor patiently waits untilthe essential bits of data arrive so that it can do its job (SN:6/16/84, p.