News & Analysis
IBM's Mira aims to beat Watson
R. Colin Johnson
2/15/2011 8:48 PM EST
PORTLAND, Ore.—While IBM's Watson supercomputer tangles with its human competitors on the quiz show Jeopardy this week, the company's high-performance computing architects will be crafting—in cooperation with the U.S. Department of Energy's Argonne National Laboratory—what they hope will become world's fastest supercomputer.
According to Top500.org, the world's fastest supercomputer is currently a 2.57-petaflop system, China's Tianhe-1a, which uses thousands of Intel Westmere processors with Nvidia Fermi graphics accelerators. In contrast, the Mira, based on the Blue Gene/Q, aims to top 10-petaflops using thousands of IBM's own Power7 cores.
This week, IBM's Watson is using 2,880 Power7 cores to compete with humans at the game Jeopardy. After the first day of the challenge Monday (Feb. 14), the three-day match stood even with the top human contestant and Watson each having 5,000 points. However, two more days of competition will decide whether a human or a supercomputer is better at Jeopardy.
The Mira supercomputer, on the other hand, will be used by Argonne National Laboratory for more practical applications, including designing electric car batteries, analyzing global climate change and simulating the evolution of the universe. IBM says that Mira will be "operational" in 2012, at which time it may become the fastest in the world.
Eventually, IBM hopes to scale up Mira toward the exascale with as many as 750,000 cores.
According to Top500.org, the world's fastest supercomputer is currently a 2.57-petaflop system, China's Tianhe-1a, which uses thousands of Intel Westmere processors with Nvidia Fermi graphics accelerators. In contrast, the Mira, based on the Blue Gene/Q, aims to top 10-petaflops using thousands of IBM's own Power7 cores.
This week, IBM's Watson is using 2,880 Power7 cores to compete with humans at the game Jeopardy. After the first day of the challenge Monday (Feb. 14), the three-day match stood even with the top human contestant and Watson each having 5,000 points. However, two more days of competition will decide whether a human or a supercomputer is better at Jeopardy.The Mira supercomputer, on the other hand, will be used by Argonne National Laboratory for more practical applications, including designing electric car batteries, analyzing global climate change and simulating the evolution of the universe. IBM says that Mira will be "operational" in 2012, at which time it may become the fastest in the world.
Eventually, IBM hopes to scale up Mira toward the exascale with as many as 750,000 cores.
|
|
|
|
|
Navigate to related information
KB3001
2/16/2011 5:16 AM EST
Reliability is going to be a major issue in the exascale era. A coordinated software/hardware appriach is needed to solve this problem.
Sign in to Reply
yalanand
2/16/2011 12:11 PM EST
@KB3001 ya reliability will be a major issue in exascale. Hopefully we will design more new tools to solve this issue.
By the way latest news is that "Watson flexed its considerable computational muscle, trouncing two all-time Jeopardy champions in last night's game and made a puzzling stumble with its answer in the final round".
source - http://www.pcmag.com/article2/0,2817,2380429,00.asp
Sign in to Reply
R_Colin_Johnson
2/16/2011 8:23 PM EST
Yes, saying that "Toronto is a U.S. City" is a bad mistake, but it just means Watson was grasping at straws. An even more telling mistake was in round one, where Watson gave exactly the same answer as the human before him. He should have known it was wrong, but did not listen to what the human said.
Sign in to Reply
Hephaestus
2/16/2011 12:42 PM EST
Watson missed the final Jeopardy question on the topic of, "U.S. Cities", when it gave the answer of, "Toronto ?????". To Watson's credit, it did put in those question marks and it bet very little (still soundly beating its human opponents). However, that was the most disappointing moment for me so far (besides the lack of a chatbot function so that it could socially interact with Alex Trebeck). It underscored that Watson is not truly intelligent but is just a highly advanced search engine with enormous raw computational abilities. The humans still win no matter what the money scores say.
Sign in to Reply
Rishiyur.Nikhil
2/16/2011 4:13 PM EST
Unless Watson was just playing a little rope-a-dope on that last question, because it was so far ahead anyway.
Sign in to Reply
goafrit
2/16/2011 7:40 PM EST
That scared me. How can IBM not know that Toronto is not in US. That is very scary.
Sign in to Reply
R_Colin_Johnson
2/16/2011 8:13 PM EST
The way IBM explained it to me, was that Watson does not scan its database, but has prefabricated indices. These are machine generated using neural-, genetic- and statistical learning algorithm so there would not be a specific list of all U.S. cities. A weird answer so obviously wrong--like Toronto as a U.S city--only means the association was roundabout.
Sign in to Reply
DrQuine
2/18/2011 6:43 PM EST
If Watson understood the category of the question (U.S. Cities), I don't understand how this error could have happened by accident. Certainly he had a very thorough atlas / map database that was used for other questions. It would be the necessary cross check before presenting an answer. Given all the question marks, I'm wondering about a sense of humor / strategy to keep the game interesting.
Sign in to Reply
dirk.bruere
2/16/2011 4:10 PM EST
The Humans will probably not win in Watson staffed call centers of the future.
Sign in to Reply
Postlethwait
2/17/2011 11:47 AM EST
I noticed a while back that in IBM domains, the .IBM comes last (supersuming countries), so maybe the USA's northern border is invisible to Watson.
Sign in to Reply
Duane Benson
2/24/2011 12:03 PM EST
I don't really think there will be any more problems with reliability at the exascale level. I'd expect it to be more reliable.
Years ago, my father described the vacuum tube computing devices he worked on as "operating on the principle of burning themselves out." I understand that some systems had technicians always on staff with the singular job of continually locating and replacing burned out tubes.
With 750,000 more or less equal processors, none of them are significant. Load leveling algorithms have been around for a long while and non-functioning units can just be removed from the processing flow via software and returned to it once repaired or replaced.
Again, hot-swapping technology is nothing new as well. That being the case, any failed units could be left in place without worry until specified maintenance time.
Sign in to Reply