Obama sets $126M for next-gen supercomputing | Hacker News Reader

Obama sets $126M for next-gen supercomputing

computerworld.com

16 pointsHusafan16 comments

Open articleView on HN

Maybe the next generation of supercomputers can figure out how to pay off the debt this funding creates.

My reaction is more like... $126m? That's it?

"IBM won a $244 million DARPA contract in November 2006 to develop a petascale supercomputer."

Similarly the prosthetic mind controlled arm took 100 million dollars to develop:

http://healthland.time.com/2011/02/11/coming-soon-the-100-mi...

Meg Whitman could have either had a sweet robotic arm or a super computer for the cost of a failed political campaign.

I remember her cost per vote acquired was something like 8$?

That kind of spending has an anti-democratic vibe to it.

I doubt it will be anywhere near an exaflop, unless you do some creative rounding. Closer to .01 exaflops.

NCSA's Blue Waters is $208m and 10 petaflops, its scheduled to go online by June 30 2011. http://en.wikipedia.org/wiki/Blue_Waters

Any idea how IP laws work under these contracts? Does IBM have the same patent rights that they would if they had developed the computer on their own?

"how to pay off the debt" -> "the next generation"

You've already figured it out, sadly.

The budget proposal is for more than $126M split between the Office of Science ($91M) and the National Nuclear Security Administration ($36M). For the first time this is called "exascale"; in the past it was known as "extreme". Advanced computing gets $465M at DOE, of which this is part. The actual numbers are all speculation as the budget is the work of Congress and not the straw man proposed by the President.

Exascale computing presents major technical problems. See the short note by Peter Kogge in IEEE Spectrum, http://spectrum.ieee.org/computing/hardware/nextgeneration-s..., and the 2008 DARPA report, http://www.er.doe.gov/ascr/Research/CS/DARPA%20exascale%20-%... .

The problems scaling traditional supercomputers into the exascale range are many and include, for example, the fact that that current computers require about 1.5 to 4 MW/PF.

The US used 61 billion kilowatt-hours of power for data centers and servers in 2006. That's 1.5 percent of all US electricity use, and it cost the companies that paid those bills more than $4.5 billion. This is expect to double by 2011. http://spectrum.ieee.org/computing/hardware/nextgeneration-s... A BOE computation says that if we were to simply scale petaflop computers to a single exascale cluster the power consumption would be 4GW, or roughly 10% of all US electricity used in the US in 2006.

Clearly computation on this scale is not a priority or the real money is elsewhere.

Exascale systems are 1,000 times more powerful than the Tianhe-1A, the Chinese supercomputer that was recently ranked as the world's fastest.

Alright, who here watched Watson on Jeopardy and now thinks what ever 1,000 times more powerful computer they build will have strong-AI?

I think very few would think that.

The heavily statistical techniques used in Watson won't turn into a strong AI just by throwing more computing power at them, any more than I can turn into a horse by walking on my hands and knees while snorting periodically.

http://www.scribd.com/doc/13863110/The-Unreasonable-Effectiv...

(Wow, that analogy ended up a little more disturbing than I'd intended.)

I love how the only justification of why the government might need such a computer, or why this spending might be a good idea, is "it will be bigger than China's".

A lot of the government's supercomputer funding is much more practical than that: its about doing simulation on nuclear weapon assemblies. During the cold war, nuclear weapons were not designed for long lifetimes or to be upgraded after construction. Most of our nuclear stockpile is or will soon be past its design lifespan. We can't just replace them with new weapons because of the Comprehensive Test Ban Treaty (yeah, it hasn't been ratified, but no administration is excited about breaking it and effectively telling everyone on Earth that the US has no interest is sticking with the NPT).

So we have weapons assemblies that are degrading every day and we need to ensure that they work (i.e., they detonate when instructed to and more importantly fail to detonate when NOT commanded to). Physical inspection and retrofitting is extremely difficult because the assemblies were not designed for that. So the only option left is massive simulation. Which is what Lawrence Livermore and the other national labs do. China's got nothing to do with it. Without supercomputers, the US lacks an effective nuclear deterrent (or at least, confidence that it has an effective deterrent).

"(or at least, confidence that it has an effective deterrent)"

Well that's the key isn't it. The only effective way to use nuclear weapons is to "not" use them. They could just as easily just release a few press-releases saying "oh yeah, we checked it out and these things totally still work", and almost as if by magic, they would continue to work.

But no one would believe that. It just wouldn't be credible. This isn't a new problem; it has been building for decades. Components age. Materials fracture. Seals crack. Everyone knows that and you can't wish that knowledge away.

I mean, the national labs that are charged with ensuring the safety and reliability of the nuclear stockpile are not some tiny 5-man operation in the middle of nowhere; we're talking about thousands of people working an hour away from San Francisco, very much plugged into the local university and tech scene.

It's not like they are publishing in detail all of their findings currently. For all we really know, they already are bluffing, and just using all the money for hookers, blow, and blackjack.

Reply on news.ycombinator.com