By John Rhea
CHELMSFORD, Mass. - The functional density of state-of-the-art computers for the most demanding civil and military computational tasks is moving up the scale to 100 billion floating point operations per second, or gigaflops, in the near term and on to a teraflop, or trillion operations, shortly after the turn of the century.
The density of high-end computer tasks reached 10 gigaflops in a single air transportable rack (ATR) measuring one cubic foot three years ago,
The projection is contained in a study completed by John Entzminger, former deputy for technology at the Defense Airborne Reconnaissance Office and a consultant to technology companies.
One of these companies is Mercury Computer Systems Inc. of Chelmsford, Mass. Barry Isenstein, Mercury`s vice president for advanced technology, says the availability of a new class of one-billion-operations-per-second digital signal processors (DSPs) will make possible what he calls a "giganode concept."
In Isenstein`s giganode concept, companies can incorporate extremely powerful commercial off-the-shelf (COTS) DSPs into new system architectures. These DSP chips, he says, will be far more powerful than the microprocessors necessary for future personal computers, he adds.
What companies have to do to implement this new architecture, Isenstein continues, is to develop the supporting interconnect architectures, software tools, and - possibly most difficult of all - cooling techniques to counter the expected levels of heat dissipation from these densities.
Mercury, whose engineers built the 10-gigaflop ATR in 1995 under sponsorship of the Defense Advanced Research Projects Agency (DARPA), will invest $100 million over the next five years in these supporting technologies, Isenstein says.
That machine, part of DARPA`s "Platinum" project, was based on a collection of 128 Intel i860 microprocessors and the supporting electronics, including power supplies, crammed into an ATR. It was used in a secret program for what Isenstein will identify only as an "airborne reconnaissance" program in which the customer was a unit of Lockheed Martin Corp.
The COTS connection, Isenstein adds, is the dual-use potential of commercial diagnostic medical imaging and military generic reconnaissance (principally radars but possibly also sonars) in what he calls an "idea diffusion" mode.
Although dual-use technologies employing COTS are usually thought of as achieving economies of scale, Isenstein says there is also a useful intellectual exchange. The military needs the greater hardware performance before commercial users do, but the commercial users tend to take the lead in software. By taking a somewhat longer view of the phased requirements, he hopes to have both hardware and software ready in time for the users.
Isenstein characterizes the idea of increasing the computational density by an order of magnitude every five years or so as "pushing the envelope [but] not off the curve."
