Slashdot Mirror
IBM Building 20 Petaflop Computer For the US Gov't
eldavojohn writes "When it's built, 'Sequoia' will outshine every super computer on the top 500 list today. The specs on this 96 rack beast are a bit hard to comprehend as it consists of 1.6 million processors and some 1.6TB of memory. That's 1.6 million processors — not cores. Its purpose? Primarily to keep track of nuclear waste & simulate explosions of nuclear munitions, but also for research into astronomy, energy, the human genome, and climate change. Hopefully the government uses this magnificent tool wisely when it gets it in 2012."
Can you imagine a Beowolf cluster of those?
Nice rack(s).
2012! Supercomputer! It's Skynet! RUN FOR YOUR LIVES!
And also to find the question that "42" answers.
There are many theories as to what this question might be, and now IBM is building a system that will solve this issue once and for all.
The simple truth is that interstellar distances will not fit into the human imagination
- Douglas Adams
I've heard about predictions of the end of the World in 2012, now I know the answaer - this machine will become a Singularity.
"The system will also act as a giant weather cock,"
Each processor gets its own megabyte of memory? Are these a bunch of refurb pcs from the late 80's?
Do you have ESP?
Because, when you put two processors on a single piece of silicon, it magically becomes one "processor" with two "cores".
Escher was the first MC and Giger invented the HR department.
My bet is that this is a typo.
1.6 PB seems more reasonable.
A group of computer scientists build the world's most powerful computer. Let us call it "HyperThought." HyperThought is massively parallel, it contains neural networks, it has teraflop speed., etc. The computer scientists give HyperThought a shakedown run. It easily computes Pi to 10000 places, and factors a 100 digit number. The scientists try find a difficult question that may stump it. Finally, one scientist exclaims: "I know!" "HyperThought," she asks "is there a God?" "There is now," replies the computer.
echo -e 'global _start\n _start:\n mov eax, 2\n int 80h\n jmp _start' > a.asm; nasm a.asm -f elf; ld a.o -o a;
flops = floating point operation per second
flop = Gigli
The article got it mostly right. It mentioned 500-teraflop once, but every other time it spelled flops correctly. Slashdot, on the other hand, fucked up the title, despite the fact that it pretty much just copied it from the article (poorly).
Hopefully the government uses this magnificent tool wisely when it gets it in 2012.
Sounds like they are going to port the quake mods to the raytrace q4 engine.
Having to work for a living is the root of all evil.
it is indeed 1.6 petabytes: http://www.eetimes.com/news/design/showArticle.jhtml?articleID=213000489
It could also be used to search for "suspicious behaviour" by searching Government databases, Credit card companies' databases, credit bureau databases, Choicepoint's, telecommunication companies' databases, airlines, and any other firm that the Government bullies into giving access.
Well, that's not as paranoid as you might think. The case against is quite simply the publicity that's been given to this behemoth of a machine, so I really don't think it's too likely in this particular case.
However this is EXACTLY how you go about putting together a machine for intelligence purposes. The key to running an intelligence service is deniability at as many levels as possible, and keeping anyone from seeing the big picture.
So you comission some huge piece of hardware, with a benign-but-complex sounding (usually simulation) function.
Then you get the low leve software put together for the platform. If that can be done in modular fashion, so much the better. You don't mix the platform with the real world data during design.
At the final stages, and presumably in house, you can write your overlaying interface (which intelligence employees will use), and only then is the pure function of the suite necessarily made apparent.
Of course there are lots of people in the design process that have a notion of how things are being put together, what they will interface with, etc, who can take a stab at the function. And sure, everyone signs an NDA just because. However, since nobody sees all of it, and the big picture is never confirmed outside of a very small number of people, and nobody is going to break an NDA to talk about part of something that MAY have some function... you essentially reduce the risk of leaking your cababilities, system spec, and intention by a not insignificant amount.
"...allowing forecasters to create local weather "events" less than one kilometer across, compared with 10 kilometers today and at speeds up to 40 times faster than current systems."
- IBM is building a computer that will be functional in about 3.5 years.
- The power of this computer, in 3.5 years, will outshine every other supercomputer currently running today.
I should hope so! What's the point of taking 3.5 years to build the thing, if it's going to be 3.5 years out of date by the time they build it?
Heck, in 3.5 years, your desktop computer will be 4 times more powerful than anything currently running today, too.
Duuh.
"City hall" in German is "Rathaus" Kinda explains a few things......
You don't need 20 petaflops to do that, you need a few tens of teraflops and a really really huge memory and really really fast IO. You'd do much better with some of the 1/4TB memory systems from Sun or IBM + spending a huge pile of money on SSDs than a real supercomputer.
The cost of the IO interconnect is a huge chunk of cash to sink into a supercomputer that you just don't need for that sort of tin foil hat application.
Slashdot Patriotism: We Support our Dupes!
Because funding for military expenditures is much easier to obtain than funding for climate research.
"City hall" in German is "Rathaus" Kinda explains a few things......
I'm sorry, I don't believe it.
I think using a BlueGene for run-of-the-mill data processing would be a horrible waste of money. There's simply no need for things like a parallel filesystem or PB of RAM or low-latency interconnects. You want to "scale out" for distributed processing like you're talking about, not "scale up".
No, I'd bet intelligence gathering is done on Google-like processor farms.
nuclear explosions whereas climate simulations don't have all the variables.
Actually I think they model the effects on decay in current nuclear weapons. Besides its not something I want them to physically test.
* Winners compare their achievements to their goals, losers compare theirs to that of others.
Interesting. There's still the matter of the software. Currently, these machines are running straightforward simulation software. I don't see weather prediction resulting in sentience any time soon. The machines will probably have to grow considerably before the more flexible software subsystems like say the load-balancing code, achieves sentience, and destroys us all.
Another reference article: http://www.eetimes.com/news/design/showArticle.jhtml?articleID=213000489 Mentions "up to" 4,096 processors per rack. So, at maximum, this would be 393,216 processors. Perhaps they are quad cores and someone took the liberty of multiplying the 393,216x4=1.6M (rounded). A more reasonable assumption may be 100,000 quad-core CPUs (400,000 cores). That would make the summarization of by only 16 times, lol.
So the real question in an immense cluster like this, is whats the MTBF?
Simon claims that the Eniac MTBF was 8 hours, although I've seen all kinds of claims on the web from minutes to days.
http://zzsimonb.blogspot.com/2006/06/mtbf-mean-time-between-failure.html
I would guess this beast will never be 100% operational at any moment of its existence.
I'm guessing the "cool" part of this won't be the bottomless pile of hardware in one room, but how they maintain this beast. Just working around one of the million CPU fans burning out is no big deal, but how do you deal with a higher level problem like one of the hundreds of network switches failing, etc?
"Science flies us to the moon. Religion flies us into buildings." - Victor Stenger
Yes, nuclear weapons have a shelf life due to the components included in them - explosives, chemicals etc.
New designs are used to maximise yield per mass, enabling you to throw a smaller warhead at a target, which means less chance of interception. It also means a smaller package to maintain, and cheaper to build, along with more warheads per unit of material.
Hopefully the government uses this magnificent tool wisely when it gets it in 2012.
SCENE: The Pentagon, 2012
Science Advisor: "President Whoever-You'll-Be, IBM has completed our 20 petaflop computer. It is awaiting your command."
President Whoever-You'll-Be: "Thank you, Advisor. We can use it to compute the long-term effects of nuclear waste disposal, weather fronts, and... just... just how much processing power is in this?"
SA: *deep sigh* "Over 1.6 million processors and a total of 1.6TB of RAM, sir."
PWYB: "My GOD, Advisor. Do you know what that much power could do? It... it could..."
SA: *another deep sigh* "It could, in theory, calculate the entire state of Wisconson to rubble. Or process the irrelevance of humanity down to a variance of 10^-24. Or, and this is what we were hoping not to worry you about, refactor..."
PWYB: (interrupting) "Refactor the planet into a singularity, yes, I know. This is a grave situation. We can only hope this much processing power doesn't fall into the hands of someone with fixed-polarity Reed-Muller expressions for incompletely specified Boolean equations and a vendetta."
SA: (long pause) "Shall I turn it on, sir?"
PWYB: "Yes, turn it on. And may God have mercy on our souls if we never need an fsck..."
Demanding constant attention will only lead to attention.
The parent said that the computer will be used for "mapping every reaction" between molecules. Presumably, since reactions tend to require quantum mechanical descriptions, I guessed the parent meant that the new computer would allow doing such calculations for all reactions in a rather large area.
I don't get what you are saying when you ask a question, relate it to the parent's post, and then say it is irrelavent.
Just a gedanken experiment to amuse myself, while noting that it actually has nothing do with simulating nuclear weapons. Don't get too worked up about it.
And, do you realize how much processing power 20 petaflops is?
Yes, it's about 2 orders of magnitude more than the supercomputer I'm using at the moment. A lot for sure, but still limited to very small system sizes for quantum mechanical calculations. At the moment, even the best methods in practice scale as N^3 or so. With my current 100 TFlops I might do a DFT calculation with O(10000) atoms or so. Two orders of magnitude more CPU power with N^3 scaling gives me roughly a factor of 5 more atoms. 50000 atoms fit into a box of roughly 10x10x10 nm (depending on the material etc., of course). Still a way to go until I'm able to do "square miles"..
If you want to go into classical molecular dynamics, then you're obviously in much better shape. With the current supercomputer that's maybe around 1E9 atoms, and since MD scales linearly, with two orders of magnitude more flops it means around 1E11 atoms. Now these fit into a box on the order of 1 um**3. Again, still quite a way to go to square miles..
In conclusion, atoms are really really tiny, and in 3 dimensions you can pack a lot of them into a very tiny volume.
Also, they so far have not needed to calculate what a nuclear bomb does for each atom (obviously, since it has been nigh impossible), and they probably won't ever need to really. You can study waves and energy effects in great detail, and simulate them accurately, without needing to know where each and every atom goes. This will simply let them be more precise and accurate, as well as speedy.
Yes, that was sort of implied in my previous post. The US nuke labs have been at the forefront in research on numerical methods in topics such as shock propagation (PPM and methods like that) and really really large FEM simulations. Obviously, the actual nuclear reactions are taken into account probabilistically rather than the full quantum mechanical treatment (as my above monologue shows, such a treatment for the primary is far beyond any computer in sight). AFAIK they use Monte Carlo neutron diffusion rather than the classical multigroup diffusion methods that AFAIK are still largely used for civilian reactor design. That being said, I'm sure they are doing a lot of atomic and quantum level simulations as well for small model systems designed to e.g. extract parameters for continuum simulations and such.