IBM’s Watson definitely has lower latency than a Jeopardy Champion

I am fascinated by the story of IBM’s Watson computer that is competing on Jeopardy. My DVR is all set for February 14-16 to see if a computer can beat the best two contestants in the history of the show. I think the last time that I watched Jeopardy was back when Ken Jennings (who is one of the two) had his big winning streak. This week, I watched the PBS show on Nova “Smartest Machine on Earth” that told the story of the development of Watson. IBM spent four years and tens of millions of dollars on this project; of course the lessons learned are much broader than just for playing on Jeopardy. The event is a showcase for IBM Research and lots of IBM technologies.

The IBM team has shared some of the stats on the product:

  • POWER7 is especially well-suited to Watson based on its parallelization capabilities and scalability.
  • The system is optimized to handle the massive number of tasks — using 90 Power 750 servers running Linux, with 2,880 POWER7 processor cores and 15 terabytes of RAM — that Watson must perform at rapid speeds to analyze complex language and deliver correct responses to the game clues.
  • Watson incorporates a number of our patented technologies for the specialized demands of processing an enormous amount of concurrent tasks and data, while analyzing information in real time.

The system is loaded with all of Wikipedia, IMDB, archives of newspapers and tens of thousands of old Jeopardy questions that helped with machine learning. The TV special showed that the system was “10 refrigerators” in size and was all self contained. That’s a lot of processing power and a lot of cores, and since I cover the networking space, was curious what switching was used. I got the following specifications back from IBM:

  • Watson is self-contained and not connected to the Internet.
  • Internally it uses a 10 GigE fiber interconnect with a line-rate switch in the middle.
  • The actual hardware is 1 IBM  J16E (EX8216) switch populated with 15 10 GbE line cards.
  • This data fabric is more than sufficient for the Watson application which is CPU and memory intensive on each node but has less than an extreme dependency on interconnect performance.

So, it’s 10 Gigabit Ethernet, not InfiniBand, and the switching technology is from IBM’s partnership with Juniper. While InfiniBand is still doing well in many HPC and Supercomputer environments, I am hearing good things from Juniper and other networking companies that are delivering 10 Gigabit Ethernet solutions that meet the very stringent performance and latency requirements such as with Watson. The computer has a ton of knowledge and can always beat the humans to the buzzer, so it hardly seems a fair fight. If you watch the Nova program, you’ll see that the computer is still learning some nuance and context for answering questions that give the people a fighting chance.

Who (or what) do you think wins? I’m a bit torn, do I root for the human team, or for the success of technology? Is IBM’s next project Skynet?