Spinoffs of Blue Gene's core technology could revolutionize a lot more than biotech.
Solving protein puzzles is sexy work for IBM, but it doesn't make for volume sales. The real payoff of the Blue Gene project will be the machine's cellular architecture - repackaged in teraflop boxes and sold to lucrative niches clamoring to do data-intensive computing.
"The big market will be at the low end of the knowledge management category," says Mark Dean, vice president of systems research at IBM's Thomas J. Watson Research Center. Financial risk analysts don't need a million processors or tens of thousands, but they might use 64.
Data capacity is cellular's big advantage. The delays involved in pushing massive models of financial markets or 3-D industrial designs across a motherboard or through a mainframe's back plane are today's biggest impediments to better simulation. By contrast, in a cellular system, processor and memory or networking components reside on the same silicon wafer, so data never needs to make the long, slow trip back across the motherboard again. Or put another way: If a current-generation supercomputer ran into an acquaintance on the street, it would look first at the guy's left eyelash, then his left eye, next his nose, a portion of his cheek, and so on. A cellular system could see the guy's face all at once.
Consequently, Blue Gene's offspring will be well suited to meet biotech's enormous pattern-matching demands. "You could run a blood sample through analysis for susceptibility to certain diseases or responsiveness to certain drugs," Dean says enthusiastically. "The key is not to take a month, but to do it before the patient walks out of the doctor's office."
Integrating memory and network components on the processor's silicon wafer will also push the price of mass-produced desktops down - way down. "We're talking about almost two orders of magnitude better price/performance than what you can buy today," Dean says - which will mean thinking differently about which applications make sense. Craig Mundie, senior VP of advanced strategies at Microsoft and former CEO of supercomputer maker Alliant, thinks that kind of power could improve communication not just between components but between computer and user. "Take the sentence-long command, 'Picture frame, picture frame, find me photos of my daughter with my mother between 1994 and '96.' What would it take to execute that?" he asks. Voice recognition for one thing, plus a search function that compares images with other images, shifts angles in 3-D, and morphs ages. With cellular architecture, your picture frame could shuffle through the family album right on its wafers without making you wait.
But even cellular computing enthusiasts admit they don't have a cure-all. Thousand-fold increases in power won't help much in areas like artificial intelligence, where we still don't know how to write a successful algorithm in the first place. And, as Mundie warns, "Think of Amdahl's law: If 90 percent of your app can run in parallel, but 10 percent can't, the processing power won't matter."
