Fresh off its win last week on Jeopardy, a “Watson”-style analytics cluster is within reach of the average enterprise, said Ian Jarman, IBM manager for Power systems, told SearchDataCenter.com.
“With Watson, one of the design goals was to create a system that could be readily used by other applications,” Jarman said, for instance the medical and legal communities. Thus, even though Watson runs sophisticated analytics software, it is based on hardware that is familiar to many data center managers, namely, 90 of IBM’s own Power 750 servers, each with 32 POWER7 cores, each with between 128 and 512GB of RAM. Those are the same Power systems frequently found in enterprise data centers running traditional OLTP applications like ERP systems.
“Some people have labeled Watson a supercomputer, but it really is not,” Jarman said. Rather, “each individual node is very efficient.”
That’s in contrast to previous IBM projects like Deep Blue, the 1997 supercomputer that beat chess master Gary Kasparov. That system was also based on POWER, but it also contained hundreds of specially designed chess chips. With its highly customized hardware configuration, “it was a one-off,” Jarman said.
But while Watson is based on “commodity” parts, Jarman did emphasize the Power system’s prowess over x86-based clusters. While Watson could run on x86-based systems, the Watson design team estimated that it would require three to four times as many nodes to achieve similar performance if it had designed Watson around Intel-based systems. “The fact is, it really was the case that the P7 and its massive on-chip memory bandwidth was critical to its performance,” he said.