Blue Skying

The ultimate business software beckons--if you can wait 10 years for a computer fast enough to run it

Imagine a 3-D computer model of the earth. Not just one of those pretty ocean-and-cloud pictures, but a virtual planet into which you could zoom to examine farmland, cities, and streets. You wouldn't be limited to physical features: you could check population densities, the distribution of wealth, the availability of goods. Best of all, the model would be dynamic: it could display the spread of buying trends, the flow of commodities, even the formation of traffic bottlenecks, all in up-to-the-minute detail.

This "Mirror World" would be the mother of all business tools--a real-time simulation of what's happening everywhere. Say, for example, you want to hone your marketing plan. "You could take data from all the companies in the United States, every transaction, down to the last bottle of beer bought at every grocery store," says Rick Stevens, director of the math and computer science division at the Argonne National Laboratory, outside Chicago. "You could find out exactly how much money was being made where and at what level of the economy. You could discover where the inefficiencies are." Then you could target your product exactly at those niches.

You could analyze stock markets to find small anomalies to use to your benefit. You could study manufacturing capacities and spot opportunities to make products at unusually low cost. You could determine which countries are most ripe for each low-cost product. And games--well, let's not even get started on games.

Talk about killer applications.

Just one problem: the computer required to run Mirror World has to perform a quadrillion floating-point operations per second, a computing speed known as a petaflops. That's a thousand times faster than the speed of the fastest-running computer today--a machine built by Intel for Sandia National Laboratories, in Albuquerque, capable of a trillion operations per second, or a teraflops. A teraflops alone is as fast as a thousand garden-variety commercial supercomputers.

The good news, however, is that petaflops computers capable of supporting Mirror World could be up and running in a decade. And though you won't be able to afford one, you probably won't need to own one. The government will most likely build it, and your business will be able to rent out its power via the Internet, on a relatively affordable basis.

That, at least, is the consensus vision of the Interagency Petaflops Initiative Computing Group, a partially federally funded collaboration of computer scientists that has begun planning a petaflops computer. The group has met four times in the past two years, enlisting the help of legendary supercomputer designers Seymour Cray (who died last October) and Burton Smith. The members of the group have come to the conclusion that such a computer can be built by around 2006. But it will require radical computer technology that won't even vaguely resemble the stuff of the silicon dinosaurs we're using today.

Trying to design a computer capable of performing a quadrillion operations per second presents some extraordinary problems. For example, imagine trying to get petaflops performance by yoking ordinary PCs into one giant computer. "It would almost fill the Empire State Building," calculates Thomas Sterling, a computer scientist at the California Institute of Technology. It would require more electrical power than could be wrung out of Big Allis, a 1-billion-watt generator that serves New York City. And it would cost somewhere between $50 billion and $100 billion.

The first challenge: no one chip can operate fast enough to do a quadrillion operations one after another in a single second. So the petaflops group figures its computer will have to do at least a million actions at once. Either it will have a million chips doing one thing at a time, or it will have about 100 extremely fast chips all capable of doing 10,000 or more operations simultaneously.

But that leads to another problem. With a set of interlinked, extremely fast chips, data don't have much time to move from one chip to another before the chips are busy with the next operation. Fortunately, data move at close to the speed of light--about 186,000 miles per second. But even at that speed, a 100-chip computer running at only 10 teraflops would have to be contained within a 3-millimeter-wide square.

Then you've got power problems. The faster and more complicated the processor, the more power it needs. A conventional machine running at a petaflops clip would eat up some 100 megawatts of power, which means its cooling requirements alone would make it the size of a three-story building--and even then it might glow in the dark.

The petaflops group figures that a way around some of the problems is a superconductor technology called rapid single-flux quantum logic (RSFQ). The technology is so new that its inventor, Konstantin Likharev of the State University of New York at Stony Brook, didn't have much time to design applications for it before the petaflops group came knocking. At the heart of the technology are microscopic rings that can instantly switch between being magnetized and being unmagnetized--representing the 0s and 1s of which all computer data are composed--while consuming virtually no power. The rings take the place of conventional transistors, leading to chips that operate at about 1,000 times the speed of today's PC chips while using hardly any more electricity. The bad news about RSFQ superconductors is that they work only at temperatures of about -451 degrees Fahrenheit. That means the computer has to be placed inside an ultra-high-tech refrigerator.

Another big gain in performance can be had through processing-in-memory (PIM) technology. In conventional computers, the processors that provide the computing horsepower are located on separate chips from the memory that holds the data to be processed. In a PIM computer, the processor is buried in the middle of the memory chips, cutting down on the time it takes information to travel between processor and memory. As a bonus, a PIM computer is relatively easy to program because data don't have to be steered between processor and memory chips.

Then comes the question of displaying the information. An application like Mirror World might churn out as many as 100 trillion bytes of information at a time--the equivalent of a few thousand hours' worth of movie film displayed all at once. The group's solution: virtual-reality interfaces, including what's known as a cave, in which the user travels inside a 3-D virtual world.

The federal government's National Science Foundation is currently spending $1 million on eight projects that explore potential approaches to the petaflops computer. They are looking not only at RSFQ and PIM but at such exotic computing technologies as chips that rewire themselves and a computer whose memory operates through laser beams. The petaflops group is betting that the government will up its ante to $400 million a year in research funding once the project gets up to speed.

All of this means that Mirror World may really be around the corner. Get your questions ready.

Gary A. Taubes (, a Boston-based science and technology writer, is a 1996 Knight Fellow at MIT.