IBM’s Watson is the powerful computer that defeated its human rivals in a Jeopardy match against Jeopardy’s two grand champions. In case you missed it, check out Watson here.
A question was raised on a mainframe discussion group about why Watson wasn’t done on the zEnterprise consisting of the z196 with the zBX. Watson runs Linux on POWER7 processors.
The obvious answer is that the zEnterprise wasn’t around when Watson was being developed. Watson on Jeopardy used 2880 POWER7 cores. Not sure you can squeeze that many cores into POWER7 blades on a zBX; the maximum zBX capacity is 112 blades. Even connecting the maximum of eight zBX boxes won’t bring you to 2880 cores. Find published z196 specs here.
Watson represents a startling achievement in its ability to respond to natural language. Alex Trebek, the host of Jeopardy, asked regular Jeopardy questions in the normal way. OK, Watson would get the questions simultaneously as text, but they contained all the convoluted syntax, puns, double entendres, and tricks that characterize Jeopardy game questions.
To make Watson competitive, it had to understand the question and come up with the right answer with sufficient confidence in less than three seconds. Watson didn’t have time to go out to the Internet searching for answers. IBM loaded it with all the information it might possible need right in memory—15TB worth, which drew on another 20TB of clustered disk storage. Then it had to pack in all the deep analytics and natural language parsing algorithms Watson needed to be competitive.
The resulting system consisted of 90 tightly integrated IBM Power 750 servers running Linux and containing 2880 POWER7 processor cores as well as the 15TB of onboard memory. The POWER7 ran at 3.55 GHz and had 500 GB per sec. on-chip bandwidth. Although the scale of Watson’s technology is impressive, even more impressive is that IBM assembled Watson from off-the-shelf commodity components. You can get the same components today should you want to whip together Watson yourself.
The commodity components are key to Watson. IBM didn’t go to all this trouble just to win a one-off trivia contest. It expects to sell configurations of Watson’s technology to solve real business problems. To do that, IBM needs Watson to be able to work its magic using commodity components.
Already IBM has targeted its first commercial application for Watson—medical diagnosis. In this case Watson replaces Dr. House, the scruffy curmudgeon genius from the popular TV show. Watson may lack the sex appeal of actor Hugh Laurie, but its boasts impressive technology specs.
IBM plans to optimize a version of Watson for various industries. For the medical industry that means loading Watson with vast and comprehensive medical knowledge from books, archives of medical journals, and the very latest research to respond to queries from doctors. This actually seems tame compared to Jeopardy, which required IBM to load up Watson with information on a seemingly endless array of subjects.
Another slam dunk for Watson should be customer support. Watson could be optimized and configured to respond to questions from a company’s existing and prospective customers or from its support agents, maybe assisted with speech to text translation if necessary. Watson certainly couldn’t do worse than some of the offshore support agents today.
Of course, Watson would have to be loaded with customer, product, and company data and maybe industry and regulatory data. Compared to Jeopardy that should be trivial. You probably could get away with far fewer cores and far less memory, in which case the z196/zBX combination might be fine if you can tune IBM’s DeepQA software to run there.
The system that won Jeopardy was far bigger than any single organization probably needs. Scaled to the needs of your organization and industry, Watson might turn out to be a promising new workload for a z196/zBX loaded with POWER7 blades.