Slashdot Mirror
Behind the Parting of IBM and Blue Waters
An anonymous reader writes "The News-Gazette has an article about the troubled Blue Waters supercomputer project, providing some new information about why IBM and the University of Illinois parted ways back in August. Quoting: 'More than three dozen changes, most suggested by IBM, would have delayed the Blue Waters project by a year ... The requested changes caused friction as early as December 2010, eight months before IBM pulled out, leaving the project to look for a new vendor for the supercomputer. Documents released under the Freedom of Information Act show Big Blue and the Big U asserting their rights in lengthy and increasingly testy, but always polite, language. In the documents, IBM suggested that if changes were not made, the project would become overly expensive.'"
I doubt IBM is well placed to do anything less expensive than anyone else...
One part of the article: "John Melchi, a senior associate director at NCSA, said last week that there is a variety of vendors available, which he compared to a choice of car dealers." Then another part: "Though she declined to answer technical questions, the FOIA documents mention clock speed as an issue."
OK, supercomputer vendorscar dealers, raise your hands: how many of you have 4+ GHz CPUs to sell? Standard, commercially available POWER7 cores run up to 4.25 GHz. That's the second highest clock speed CPU in the world, and by a considerable margin. (The highest in the world? this one, at 5.2 GHz.)
Could it be that academics demanded their idea of perfection and were unsatisfied with mere best available reality? That's never happened before.
IBM told you to take your $300 million project somewhere else? If that doesn't say VOLUMES about your project management/specification process, I don't know what does.
Something tells me that the only reason this article exists is because of the intended pun in the title.
It may be that other vendors appear to be able to do the project for less money per TFLOP. If (example with made up numbers) an SC based on the POWER7 cores has 100,000 cores and they cost $1000 per core, but another Intel-based SC with 200,000 cores can do the same work and costs $400 per core with the same operating cost, then the latter machine is cheaper for the performance required - which is the figure of merit.
It's easier to be a result of the past, but more fun to be a cause of the future! http://www.spacefinancegroup.com/
Illinois is broke as all hell. They're setting themselves up to be the next California. I moved out around a year ago before they raised state income tax. At that time they hadn't paid any of the in state colleges what they were owed for around 18 months.
U of I also came out with a genius plan to 'lock in' your rate for 4 years, so if there is a short fall, the next year is going to have a huge jump in tuition. It's getting to the point where in-state for U of I is as expensive as out of state if you were to go to Purdue, Michigan, UW Madison, etc. Out of state tuition is up near the cost of private schools.
Sorry. I guess my title was just as full of shit as TFA's You can read the article to confirm but basically it says, "we can't give out the facts".
Having to work for a living is the root of all evil.
Wow, how far that company has come; from WWII-era hollerith cards to supercomputers.
Just as the original cards were used, I hope IBM uses its resources for good purposes and to further mankind's knowledge.
LOL
... and they all spinned and lied like used car salesmen.
It would be apt to rename it into Deep Waters now.
Such projects are always based on estimated performance numbers. It looks to me like the estimated (and contractually signed) target performance was higher than what IBM could deliver in the budget envelope and the target timeframe. Probably the technology advance was not delivering as expected. As the U of Illinois was not ready to soften on some aspects to make it fit IBM had the choice of either delivering much more hardware at a loss or to pull out.
From my (limited) search it looks like the project was signed in 2008, so back then IBM estimated they can deliver a PetaFLOP for $200M in 2011. It looks like they were wrong.
I witnessed similar situation where the machine could deliver the promised performance using a benchmarking program, but real apps were unable to get to similar numbers. Improving the software stack made the performance available to apps, but it took 2 years to get there. The hardware was performing well, but immature software was spoiling it.
Markus