Talking to a friend at IBM last night, I learned that Watson, technically speaking, is not a supercomputer. This was a bit disconcerting to me, since I refer to it as one a couple times in the book (which comes out as an ebook on Wednesday).
Without getting into a long discussion of MIPs and Petaflops, two points: As I understand it, broadly speaking there are two types of supercomputers. The traditional kind, which is important for jobs like modeling the bending of proteins and the blasts of atomic bombs, requires insane amounts of mathematical calculations. Watson is definitely not one of those.
The second kind is more like the Google computer. It’s called “…data-intensive supercomputing….” In Google’s case, it involves clusters of commodity computers working in concert to process the chaos of unstructured data that you find on the Web. Watson is closer to this model. Its specialty is words. (Here’s a 2007 paper on it by Randall Bryant, head of the computer science dept at Carnegie Mellon. That paper started me on a path that led to a BusinessWeek cover story I wrote on Google’s cloud.)
The Watson you’ll see playing Jeopardy on Feb 14, 15, and 16 runs on a cluster of IBM Power 7 servers, and it features 2,880 processing cores. I guess by today’s standards, that unit can’t handle enough calculations per second to qualify as a supercomputer. But I think there’s also a bit of marketing behind IBM’s word choice. The company wants its customers to see Watson machines as business computers. So they don’t want to associate Watson, in their customers’ minds, with the Sandia National Laboratories or the National Weather Service. These things should be affordable.
Of course, the other point about Watson is that it’s really just a software program. It could run on all kinds of different computers. In fact, the Jeopardy team at IBM has actually built two Watsons. A slower and more forgiving one is for development. Its speedy twin plays the game.
