
Skynet's coming
nuff said
Famous US military crazytech agency DARPA has issued a challenge to the IT community: do you think it's possible to build a petaflop supercomputer that fits in a single air-cooled 19 inch cabinet and requires no more than 57 kilowatts of power? One which requires no special programming skills to use? The challenge is laid out …
A rise of the machines story?
Abtraction from the hardware via a self aware OS?
Then there's the generator, perhaps they might think it should be a small nuclear reactor.
I hope I live to be an old man and die before that happens.
Perhaps the film Eagle Eye is very prophetic indeed and then of course there's I, Robot.
For goodness sake, don't give these nutcases positive feedback or we're all doomed!!! :)
Skull and crossbones because it'll be only a matter of time before it considers us meatsacks surplus to requirement.
Computing will get there given time - given today's server power and Moore's Law, I'd be interested to see a calculated guess at when DARPA's requirements will happen naturally.
However, given the more rapid advancement in communications, I wonder if the answer in the meantime won't be to have a high powered, compact, hot brain in every automated vehicle, but to outsource the brain to a data centre, using high-bandwidth wireless communications (which I'm sure the military are more than capable of acquiring).
These extra icons just make it more difficult to choose. And what's with the lack of dithering on the thumbs up and sad face?
I wonder if that would be "a New Virtualised Operating System with Irregular and Unconventional Networks InterNetworking Joint Adventures and Special Application Programs" they would be looking for?
The one that is lightly touched upon/hinted at in "Use DPI Better, Now …… alsjeblieft. Which probably means buying in SMEs*" ... Posted Friday 26th June 2009 07:31 GMT .... http://www.theregister.co.uk/2009/06/25/uk_cyber_security_strategy/comments/
It certainly meets all the core requirements and that would make it very valuable, and to rivals into the same dominance posturing/hyperbole, priceless.
You know, whenever I meet a young aspiring programmer, I tell them "get into AI - it's a job for life".
Lol, I can already give you the core code for Darpa's "self-aware OS". It looks like this:
switch (machinestate) {
case (state1)
dostate1();
break;
case (state2)
dostate2();
break;
...
...
default:
do_something_sensible_given_the_circumstances();
}
Now, if DARPA can just get one of their bright youg things to write the simple do_something_sensible_given_the_circumstances() function, I will happily code the rest of their OS for free.
Bottled water, anyone?
I've seen some of the top 100 supercomputer clusters, if all you want is raw CPU its not that big (excluding the aircon and transformer bits). But your disk storage, all those spinning disks to give you the petabyte alongside the petaflop -that's what takes up space.
not possible. Can't have.
Most especially not if you don't need a deep knowledge of the architecture.
Must try harder.
Fer pete's sake do the sums, 1 petaflop = 10^15 flop.
1 nanosecond = 10^-9 seconds. About the time taken for a photon in vacuum to travel 1 foot.
So, about 10^6 calculations per nanosecond, meaning about the time it takes for a ray of light to cross your box (being very generous, cos it's going to be bigger than a foot cube). Comms alone is going to be a problem, innit.
Either some William Gibsonesque quantum crap (right now I've no good reason to believe quantum stuff will ever work in a big way), or a shift to something else (analogue computing anyone? entailing a severe 'paradigm' shift, and with results not measurable in flops anyway), or something I haven't thought of.
Grow out of it, darpa. The real world awaits.
Ew. There's no meaningful comparison between organic brains and computers, because they're apples and oranges. I mean, any pocket calculator can do a good number of tasks much faster than my brain.
And intelligence isn't equal to number of neurons anyway. Simulate a neural network with as many nodes as neurons in a mouse's brain doesn't make a virtual mouse. And simulating a neural network with as many nodes as neurons in a human brain won't make an AI.
"AI doesn't exist." ... By Mage Posted Friday 26th June 2009 14:53 GMT
Thanks, Mage. Such delusional nonsense permits such covert development as you would obviously not believe.
And that is not all, for there is also the following ...<<< And ITs Immaculate Stealth comes in Third Party Disbelief of such Stated Facts prefering as so many can and do so easily do, to dismiss such Tales as a Manic Fiction rather than Indisputable Shared Truth. But although Freely Available to Everybody and Anybody, IT does require QuITe a Lot of Personal XXXXPerience to Master NEUKlearer HyperRadioProActivity for Polyamoral Ubiquitous Programs ………with Promise in Deep Capture Projects.>>> which is AIMove in Parallel with .... "Shouldn't this be.... A rise of the machines story? Abtraction from the hardware via a self aware OS? Then there's the generator, perhaps they might think it should be a small nuclear reactor." ..... By pctechxp Posted Friday 26th June 2009 13:49 GMT
"Bet the.... Aliens could do it!" ... By Geeks and Lies Posted Friday 26th June 2009 14:57 GMT
In their sleep, when IT is so easy, Geeks and Lies.
Greg Fleming wrote:
"["fail" graphic] And here's the reason why:
http://www.rebelscience.org/Cosas/Reliability.htm"
Abstracted from that webpage:
[my proposal] will not only result in an improvement of several orders of magnitude in productivity, but also in programs that are guaranteed free of defects, regardless of their complexity.
Sorry, I stopped reading after that. Did anyone get further?
Author says: "If DARPA get their way, this sort of computing power will now be available in a single cabinet. Any decent-size data centre will in future be more intelligent than its human admin"
The accumulated weight of a man in pork meat doesn't make the meat smart now, does it?
A Radeon graphics card does 1.2 Teraflops for 160w of power, so 800 of them is a Petaflop for 130Kw. A 50% increase in processing power and 30% decrease in energy should just about do it.
You just need about 100 of these initially then leave them for a year or two and they'll write their own operating system.
Here's someone suggesting a 4 petaflops machine using current generation video cards: http://helmer3.sfe.se/
It doesn't quite fit in a 19" rack and uses twice as much power per PFLOP as DARPA wants, but it's not all that far off. Since DARPA does not expect to have its petaflop cabinet until 2017, it should be perfectly doable on the hardware side.
I think the programming/execution model and the smart OS may be more of a challenge than the hardware.
So that was what scuppered the Japanese 5th generation project? Not enough FLOPS.
And of course the "any ordinary language" bit optiminally spread across all the processors.
Given that most software for this lot will be written in FORTRAN/C/C++ and possibly LISP
Optimal splitting of software *without* the software being aware of the hardware structure.
Sounds familar. I think several projects have attempted this (with varying degress of success)
Or perhaps stuff them through a FORTRAN/C/C++ 2Occam converter and run that as a collection of (I dunno) parellel sequential processes perhaps?
Mine's the one with a CAR Hoare book in the pocket.
"I sure DARPA can get my location from the reg if they are really interested." .... By Mike 61 Posted Friday 26th June 2009 15:35 GMT
Mike 61,
The more I learn, the more that I may be easily convinced that they can do a lot less than is imagined. It is though a convenient fiction pimping a certain perception which imparts a sort of remote control.
That web site that's being quoted. I wonder if it realized that the page seems to overlook a rather important piece of computing problem-solving that Turing's machines demonstrate: the Halting Problem, or the theorem that it is impossible for a "computer" to determine in advance whether or not a given program will halt or run infinitely (the proof is by contradiction).
Equivalent human cognitive capability in computers may require more than algorithms and operations per second. It may require the coordinated superposition and decoherence of key elements- such as electrons in the microtubules of tubulin proteins as theorized in Orch-OR by Penrose and Hameroff.
Can this set of conditions that exist in human neurons be met in quantum computing software/hardware? If human consciousness is an emergent property of this set of conditions, what emergent properties will come from such a computer. Will it be Skynet from "Terminator" or the lifeform from "ST:TNG Episode- Emergence" or something much more 'interesting'.
The answer is asynchronous hardware.
Given their stated requriement its 20pW per operation.
Modern processors are reported to need c50% of their transistor count for clock drivers & de-skew.
IIRC worst case overhead for asynch is 25% max. Usually less than this.
Of course this means it's *very* difficult to split a product line by clock speed (ther isn't one).
But to go the whole hog you'll be looking at the whole deal. Chips face down, multi-chip modules etc.
Note that raw power is (in principal) fairly easy. The whole ARM chipset (32 bit processor, MMU, maths chip) ran about 100k gates.
Keeping it fed (and harvesting the results) is likely much harder.
But that will be *nothing* compared to the algorythmic slice and dice software and "Self Aware" do-what-i-mean OS.
Just my tuppence.
This post has been deleted by its author
I tend to hesitate to say something is impossible when referring to computers, given the staying power of stupid predictions (like Bill G's proclamation that 640KB is enough memory) but a petaflop in a cabinet is pretty unlikely in the foreseeable future. Shrinking current tech to the proper size is certainly possible but then the whole power/cooling thing becomes a huge problem as eventually you get to a point where Moore's law slows down and you can't fit enough electrons through your tiny traces to run your tiny transistors without literally melting the thing into oblivion.
They're not looking for AI they're looking for patterns and exceptions, deviations and spikes. It doesn't have to be Human Intelligence it has to be Artificial intelligence and that fucking Turing test guff is a red herring. I don't want to chat to the damn thing I want it to find things and notice stuff.
"A Radeon graphics card does 1.2 Teraflops for 160w of power, so 800 of them is a Petaflop for 130Kw. A 50% increase in processing power and 30% decrease in energy should just about do it."
And who knows, the resulting assemblage of silicon might be able to run Crysis/Vista/A.N.Other renowned resource hog[*] with all settings at max.
I know, I know - I'm already out the door ...
[*] - delete as appropriate
I hate that term "CrazyTech" - yeah, just slate those who dare to dream, call them crazy. Don't bother applying for jobs with DARPA, you're going to be labelled nuts!
Well, I beg to differ. If you don't try, you don't achieve.
Sooner or later some team somewhere (probably our Jap chums) is going to evolve a "proper" AI. Impossible? How so? The humain brain is just laying there waiting to be reverse-engineered! Unless of course you're one of those bible-bashing nutjobs who thinks the "soul" drives us.
It is possible, maybe, one day, we will create a machine smart enough to create another machine, which would be even smarter... etc and bootstrap the creation of something pretty interesting.
Whether we have anything as benevolent as a Iain M Bank's 'minds' or something more akin to Skynet, lol, is an interesting question.
If somebody makes one, how long before somebody else buys a roomful and builds a DBHPC* cluster of 'em, chewing megawatts of power and producing oodleflops of grunt?
Alright, two questions. How long after it's done before DARPA tout for *that* in a 19" cabinet......
*Dog's Bollocks High Performance Computer.
You described exactly what DARPA does. They put out a bunch of crazy specs and see if any commercial firm (or individual) can come up with a viable proposal. If someone can come up with a proposal they give you a sweet job and let you try and build it - doesn't matter if you make it work or not, you still get paid and get a good line on your resume. There are no non-bureaucratic "careers" at DARPA, you get hired to work on a project then your time is up when the project is over.
As an interesting footnote, often people who get hired by DARPA and successfully complete their project start companies immediately after their termination from DARPA to produce those projects for various govt. agencies. Pretty good work if you can get it. Check out their website sometime and look at the jobs section...