DARPA Wants a 19" Super-Efficient Supercomputer 200
coondoggie writes "If you can squish all the processing power of, say, an IBM Roadrunner supercomputer inside a 19-inch box and make it run on about 60 kilowatts of electricity,
the government wants to talk to you. The extreme scientists at the Defense Advanced Research Projects Agency this week issued a call for research that might develop a super-small, super-efficient super beast of a computer. Specifically, DARPA's desires for Ubiquitous High Performance Computing (UHPC) will require a new system-wide technology approach including hardware and software co-design to minimize energy dissipation per operation and maximize energy efficiency, with a 50GFLOPS per watt goal."
These aren't normal scientists... (Score:5, Funny)
57KW air-cooled 19" Rack? (Score:2)
You must be joking. That's like packing in 30 2KW electric fan heaters into a rack, obstructing the airflow with a ton of other junk and praying it won't melt. Good luck with that.
Re: (Score:2)
A KW of heating power is a lot different from a KW of processing power I imagine. Especially if some of that power is for refridgeration.
Re: (Score:2)
Expended energy always becomes waste heat. There's not much difference between drawing 500 watts in a space heater and drawing it with some other appliance from a heating perspective unless the appliance absorbs energy in some way (such as boiling water).
Even with a fridge.
Re: (Score:2)
Yes, but we're not talking about a room, we're talking about the confined space of a rack. Nothing was said about not being able to produce heat, as long as it gets out of the case before the hardware fries.
Re: (Score:2)
Never mind actually, someone in another thread has pointed out the obvious: that this is a rack, not 1u of a rack :) 57kw in a rack is probably quite doable.
Re: (Score:2)
Sure it is, just shove 42 1,500 Watt 1U power supplies in a standard cabinet.
Re: (Score:2)
Re: (Score:2)
I think it is doable. The whole machine must be air-cooled but nothing in the RFP says that liquid cooling could not be used internally.
Re: (Score:2)
Exactly, and you could use something with a high density like molten lead in titanium pipes to move the heat off the chips. And by chips I mean a single piece of silicon 19" wide and 24" deep. That's going to be the only way to do it with today's tech.. I mean, according to this [ibm.com], the PowerXCell 8i die size is 212mm^2 and they do about 100GFLOPS each. So to hit 1PFLOP, which is 1000000GFLOPS, you'd need 10000x212mm or 212000cm^2 or 2120m^2 of silicon. The aforementioned 19x24" chips would be 2207cm^2 ea
Re: (Score:2)
Incidentally, 960KW = about 909 btu's per second. 1 ton of refrigeration can move 3.333 btu per second meaning you would need 273 tons of cooling for the rack just to keep up.
Re: (Score:2)
> Incidentally, 960KW = about 909 btu's per second
Where did you get 960kW?
Re: (Score:2)
No need for such extreme measures.
Re: (Score:2)
You know... all they have to do is wait 10 years and they can just pick up a few Playstation 5's or Xbox 4's from Wal-Mart to do the job. In case they didn't realize, all those nifty bullet-points are highly desirable things for computers in both the consumer and commercial markets as well:
* Highly efficient energy usage
* New systems and programming models to develop for massively concurrent processing
* Highly fault-tolerant
Advanced computer technology is something that the market is pushing ahead at a phe
Re: (Score:2)
IANAP, but... how much of that 15kw is used locally in processing though, versus transmitted as energy/signals to another network node?
Re: (Score:2)
99.99...%
Re: (Score:3, Insightful)
Exactly. There are always incredulous responses to this kind of challenge. Everything is impossible. Until it's not anymore. That's research, that's progress. There's no better way to get people to innovate on crazy shit then to tell them it's almost impossible.
Could at least editors have a look at TFA? (Score:5, Informative)
It's a 19" _rack_, not _box_. As in, the standard (non-telco) datacenter rack size, accomodating up to 42U, 19" wide.
Re: (Score:2, Funny)
19" tower and space elevator in one (Score:5, Funny)
Two! Two! Two projects in one!
Re: (Score:2)
So seven projects then?
Re: (Score:3, Informative)
How tall does this rack have to be? :-) ) so this makes the standard rack able to contain a little over 6 feet worth of hardware, or 185ish cm. Of course the rack itself is usually a bit taller since it has a base and some fans on top (let me stress: usually).
Typically 42 Rack Units, as I said in the original post. A rack unit is 1.75" (and we use them even here in Europe so at least servers fit in racks, fortunately
Re: (Score:3, Insightful)
hmmm... fits in one rack and has enough processing to do word recognition on all of the calls coming in to one telephone central office simultaneously. I wonder what they want a whole bunch of these for?
Re: (Score:2)
Re: (Score:2)
Naw, how deep can I make the cabinet ;^)
No problem (Score:5, Funny)
Just stick a human brain in a bucket. It's small, quiet, cool and just feed it a Cheeto every once in a while to keep it running.
Re: (Score:3, Funny)
Well, since Darpa also wants the operating system to be self aware, that's a start.
Re:No problem (Score:5, Interesting)
Just stick a human brain in a bucket. It's small, quiet, cool and just feed it a Cheeto every once in a while to keep it running.
And since the human brain has a computational power of 100 petaflops at 20 watts [movementarian.com], it'd well exceed DARPA's requirements.
Re:No problem (Score:5, Funny)
And since the human brain has a computational power of 100 petaflops at 20 watts, it'd well exceed DARPA's requirements.
Only before you subtract the power used to think about pr0n. Or you can use a female brain, it'll no nuclear on you only once a month.
Re: (Score:3, Funny)
What happens if we use an Ab Normal brain?
Re: (Score:2)
Uhhhmm, removing the female brain from the female body precludes the nooclear monthly. Or, is that the monthly nooclear? Whatever - no hormones, no nookie.
Re: (Score:2, Insightful)
It's interesting how the "estimated computational power of the brain" seems to increase every few years to keep it very far ahead of the fastest supercomputer of the time.
Re: (Score:2)
Unless we are talking about these [wikia.com] who an earth would like to stand in 19" rack for the whole business day!
Zombie bait (Score:2)
Just stick a human brain in a bucket.
Yeah, we tried that already . . . but brains in buckets tend to attract too many zombies . . . you end up spending way to much money on ammunition for gun shots to the zombie heads . . . though, the sysops seem to love the action.
Going against Resistance is futile. Sidestep it. (Score:3, Interesting)
Re: (Score:2)
Which only shifts the heat to somewhere else, by using power, creating even more heat.
I don't think that is what they want. ^^
60,000 watts? (Score:2)
Is that all they're allowing? Power nazis.
Where will be that cabinet? (Score:2)
If i could do that (Score:2)
Do you think id admit it and have the Feds take it from me for nothing and classify it? No thanks.
And for the record, it wouldn't be that hard to do, as long as you wanted a semi-dedicated supercomputer and not a general purpose box.. But no, i wont tell you how, even if i was authorized.
Thermionics and stuff (Score:2)
Pack thermionic converters between the components. They'll help cool and recover some power from heat back to power. They can be on the board, or placed on a cover over it in such a way as to fit between the board components. http://web.mit.edu/newsoffice/2001/electricity-1205.html [mit.edu]
Build in parallel processing with 16 processors, 4 on each side of a 4D-cube, as in the Connection Machine http://en.wikipedia.org/wiki/Connection_Machine [wikipedia.org]
Three boards, stacked. Top, thermionics on the underside fitting between the
Re: (Score:2)
thermionics or thermoelectrics: either cool or recover power. you won't do both. And you'll add more net heat in the box if used for cooling. If you can keep that heat away from critical components, there may be a net benefit. If you use it to recover power, it'll have to be in the thermal path and will increase the thermal bottleneck in getting heat out. So you're recovered power comes at the cost of hotter heat generating components.
NVIDIA (Score:2)
Re: (Score:3, Interesting)
I imagine they will build something along those lines. Lots of highly specialised cores that can do Floating Point really well if it carefully compiled for them; some switches for some fast short-range network protocol probably and a few general purpose cores to manage things. Maybe some field-programmable components so that you can customise the hardware for new applications. The current nVidia Tesla series achieves around 1GFLOP per Watt, and you can get 1 TFLOP, consuming 1 KW per U, (ignoring host proce
Re: (Score:2)
Careful Compilation?
â Develop new technologies and execution models that do not require application programmers to explicitly manage system complexity, in terms of architectural attributes with respect to data locality and concurrency, to achieve their performance and time to solution goals - programmability.
Billion way parallelism is also mentioned. So. It has to be easy to program
Re: (Score:2)
You're missing the point-- Darpa feels that it is ill served by current commodity supercomputers, and wants something revolutionary. The deadline for delivery is in 2017, so it's unlikely that today's tech comes close.
Comment removed (Score:3, Funny)
Will fit inside your Car Analogy (Score:5, Insightful)
Having such great computational power available to every single vehicle would open up a huge realm of possibilities: Combine it with sensors you could detect damage and minimize its effects by comparing the vehicle's response to a detailed finite element model. You could do on the fly aerodynamic analysis, allowing a fighter to keep performing to it's best even after damage has significantly altered it's shape. You could manage the control of thousands of actuators, allowing you to create a shapeshifting walker out of programmable matter [wikipedia.org], and you could definitely do learning/optimization algorithms that would allow for an AI capable of a significant amount of learning. Combine this with the amount of image processing it could do, and you're very near a completely autonomous, smart enough combat vehicle.
While it's a too big for a man portable system, with work, you could fit such a device (and a power source) into something as small as a motorcycle or a somewhat scaled up iRobot Warrior [irobot.com]. That's not much more than man sized. It may not be a T-800, that much computation in that small size and power envelope is enough build a near-man sized autonomous fighting vehicle that can see, learn and adapt with an endurance on gas of several hours. It's a bit frightening to consider.
--sabre86
Re: (Score:3, Informative)
But where, exactly, would the batteries that can push 60 kilowatts go? I don't think they would fit in the trunk of a Mazda Miata with this magical imaginary computer.
Or more importantly, batteries that can push 60 kW for any period of time. I think that with enough cells, which you can make about as small as you want, you might get the power, but you definitely won't have the energy to run it for anytime whatsoever. The energy density is nowhere near good enough. But, Sticking with the Miata example, there's easily enough power under the hood to drive both the car and the computer, particularly with a high output option like the BPT [wikipedia.org]. You just need a generator, like the
Re: (Score:2)
I have a battery rated for 62.5A @11.1V. I could easily fit 100 of them within the trunk of my Mazda Miata, with lots of room to spare, which gets you about 70kW. Granted, for only two minutes, but in fact a few hundred could fit in the Miata...
(the battery is 16mmx44mmx141mm and weighs about 200g. )
Alternate way to get one (Score:3, Funny)
aka (Score:2)
playstation 4
I have one (Score:2)
BEOWULF! (Score:2, Funny)
Why? (Score:2)
Why would DARPA want this? Maybe they want a AI that can navigate aircraft or gound vehicles? BTW, I think it's ironic that autonomous operation seems easier to develop for aircraft than for ground vehicles when you consider that pilots get way more respect than the average municipal bus driver.
BTM
Hmmmm (Score:5, Funny)
So now we know what the hardware requirements for Windows 10 are going to be.
Re: (Score:2)
Re: (Score:2)
Government job (Score:2, Funny)
Given this is the government would I still get funding if I developed a computer that was capable of 50 Gwatts per FLOP?
only 60KW? (Score:2)
get rich or die trying!
The company who were closest have gone broke (Score:4, Interesting)
A firm called SiCortex was selling just this sort of compact, energy-efficient supercomputer. They shut down a few weeks ago because an investor pulled out.
It's a damn shame, they had really cool stuff. If I was Johnathon Schwartz I wouldn't have pissed away $1 billion on MySQL (it was worth maybe $10 and a stick of gum), I would have been out the front of SiCortex banging on the door with a chequebook.
Oh well.
Re: (Score:2)
The SiCortex website mentions their top of the line machine using about 900mW per 1.4GFlop processor (1.55Glop/W), but the overall system uses 20KW for a 5832 processor system, giving a system figure of (5832 * 1.4) / 20,000 = 0.40GFlop/W.
http://www.pathscale.com/products/high_capability_system_sc5832 [pathscale.com]
Now, consider that DARPA wants 50GFlop/W - a factor of 100x improvement over the SiCortex number.
So, I'd hardly classify SiCortex's products, cool as they may be, as "just this sort" of energy-efficient superco
Not one, the other (Score:2)
Forget about computing FLOPS and instead first consider this a thermodynamic problem.
Solve the thermodynamics, find a little space for CPU cores, solved.
It's doable, but too expensive (Score:5, Informative)
This is actually probably quite doable, but would be filthy expensive.
Most people don't realize, but digital electronics is way, WAY ahead of what you get in your home PC, if you're willing to pony up the cash.
For example, non-Silicon based semiconductors often outperform the good old standard stuff significantly. Silicon is by no means the fastest, it's just the cheapest. Gallium Arsenide and Indium-based materials can both clock many gigahertz higher than Silicion for the same process size and power dissipation. They're toxic, fragile, and the largest wafer sizes are tiny, so not exactly mainstream, but available now.
The real performance king though is the Rapid Single Flux Quantum [wikipedia.org] process, which can go over 100 GHz easily. It's used in things like radio telescope amplifiers and high-performance DSPs for military radar. Sure, it requires liquid helium cooling, but it also only requires milliwatts per gigaflop, so it's just about the only technology that'll let you squeeze a petaflop into a box and not have it melt into slag. That still means you'd need something like a kilowatt of cryogenic cooling, which is nontrivial, but still, I'd say it's doable with a bit of engineering wizardry.
50 GFLOPS per Watt = Hard (Score:2)
Just for some reference, an Intel Core 2 Quad can throw down about 50 gflops, but it runs on 95-130 Watts of power. To even come close to this 50GFLOPW number, you'd be designing your own HPC chips for sure.
http://www.pctechguide.com/26quadCore.htm [pctechguide.com]
Re: (Score:3, Informative)
19 inch box?
The IBM Roadrunner:
"occupies approximately 6,000 square feet..."
http://en.wikipedia.org/wiki/IBM_Roadrunner [wikipedia.org]
Good luck with that...
Re: (Score:3, Insightful)
17" tower? 3.8 GHz?
I'm sure the thinkers of 1941 would be shocked to know what we can do now, given they were running 10 Hz on this:
http://en.wikipedia.org/wiki/Z3_(computer) [wikipedia.org]
Simpler solution. (Score:2, Insightful)
Redefine a Gigaflop. Say 1 billion floating point instructions per century.
Hey - It worked for hard disk manufacturers for gigabytes.
It works all the time for food companies when they say something now has "only" X number of calories per portion, by making the portions something like "2 potato chips."
It works for ISPs for "unlimited Internet access".
It works for Microsoft for "most secure [insert whatever] ever."
It worked for George "Mission Accomplished" Bush. Kinda ...
It'll work for Barack "N
Re: (Score:3, Informative)
Redefine a Gigaflop. Say 1 billion floating point instructions per century.
Gigaflop doesn't even have a time dimension.
Re: (Score:2, Troll)
Re: (Score:2)
Yes, I am sure. You should have noted the lack of an 's' at the end of the word "gigaflop."
I can't believe two responses have been modded up for such blatant illiteracy.
Re: (Score:2)
They want 50GFLOPS/w so you start talking about GFLOP people assumed it's a typo.
PS: Hand them something that does 50GFLOP per year / w and they will say it's out of spec because they said 50GFLOPS/W.
Re:Simpler solution. (Score:5, Informative)
Are you on drugs? Sure it does: FLoating point Operations Per Second.
Hint - they're looking for a machine that can do 50 gigaFLOPs. Such performance is always measured per unit of time. Same as 1 horsepower is 550 foot-pounds per second.
If you google for it yourself, you can keep your beginners-level trainee deck swab geek card :-)
Re:Simpler solution. (Score:4, Informative)
That's right -- gigaflops has a time dimension.
Gigaflop, on the other hand, doesn't. One gigaflop is a billion floating-point operations. One gigaflops is a billion floating-point operations per second. Contrary to "obvious" rules of grammar, the "s" isn't pluaralization, it's the unit "seconds".
Re: (Score:2)
Re: (Score:2)
It doesn't get any better if you write out the full term "kibibits". Thankfully, I've never had to say "gibibits".
Re: (Score:2)
If you google for it yourself, you can keep your beginners-level trainee deck swab geek card :-)
Wow. Two uber-confident and completely incorrect responses in a row now, both citing the exact same text that should have been made it completely clear that without an 'S' in GigaFLOP there is no seconds.
Re: (Score:2, Interesting)
FLoating point Operations Per Second
FLoating point OPerations
Neither are really good acronyms, (That would be FPOPS and FPO), but they are the accepted terms.
HOWEVER: It is all nitpicky geek-out, I'm-better-than-you, You're-So-Dumb-You-forgot-the-"S". B.S.
Please.
This is Slashdot, not middle school.
hard to tell some times, I know.
Re: (Score:3, Informative)
Supercomputer performance is always measured in FLOPs per second.
Sorry, as a guy who has built supercomputers (not the piddly CoW type either) and now consults on them on a daily basis, I can tell you that's not true. Common, but not an "always" worthy of bolding - dhrystones and whetstones and their modern versions in SPECint, SPECfp and their _scale counterparts and mcalpin's STREAM benchmark which reports bytes/s in addition to flops. Not every FLOP is created equal which is why more sophisticated measurements exist.
In fact, it is precisely because of my experience
Re: (Score:2, Interesting)
It's totally unscientific, but I just ran Xbench on my 2009 Mac mini and got around 3 GigaFLOPS. It's not very accurate, but probably good to within an order of magnitude.
According to Wikipedia's supercomputer [wikipedia.org] article, that compares roughly with a 1985 Cray-2, which cost about $25 million at the time and was the size of a large closet.
All we have to do is wait about 25 years.
Re:Yeah sure (Score:5, Funny)
19 inch box?
They didn't say how high.
In other news, progress on a space elevator has been confirmed. Curiously, it's 19 inch wide.
Re: (Score:2)
The IBM Roadrunner:
444.94 megaflops per watt
So... to meet the DARPA goal... you'd only have to be more than 100 times as efficient.
I don't think it's going to happen this year
Re: (Score:2, Insightful)
Defense Advanced Research Programs Agency
These are the guys who fund the crazy stuff, like robotic exoskeletons (Starship Troopers), Electronic Telepathy (via radio/net), and more.
NOTHING they fund is expected to bear fruit quickly, but when it does bear fruit, that fruit is like gold (case in point, a little thing once called ARPANET.
Re: (Score:2)
Re: (Score:3, Interesting)
Yeah, my first thought on this was whether perhaps those were the requirements to get the things inside every AT&T-style NSA listening room.
Re: (Score:2)
So, in your opinion, the entire government is devoted to unloving and unprotecting your freediom, eh? How about the NSTA, the fellows who investigate transportation accidents? They are taking away your right to freely die due to a repeat error. How about the Coast Guard? They'll be taking away your ability to completely die due to operator stupidity. Errr...NSF...taking away your rights to no money for science. NIH...clearly out to fund medical research allowing you to die without a remedy to stop it. FDA.
Re: (Score:2, Interesting)
The defense department tends to be the place that these people are most concentrated: the same Defense Department that DARPA serves.
I doubt that ANYONE in the government does not Love and wish to protect Freedom:
There are those, however, who may undermine it unintentionally due to a lack of understanding of what their "payment to supports," will actually cause in the long
Re: (Score:3, Funny)
Back *waaaaay* off, man. I'm an *extreme* scientist!
That sounds like a nice bumper sticker. For the rear bumper.
Re:extreme scientists (Score:5, Funny)
Back *waaaaay* off, man. I'm an *extreme* scientist!
Oh, yeah? Where's your badge [scq.ubc.ca]?
Re: (Score:2)
Lots of funny ones there, but I think Hannibal Lector did the captions:
The "I've eaten what I study" badge.
Recipients have prepared their object of study as a cuisine item for eating. Hopefully, the minority of MDs are ineligible for this one. (J)
Re:oblig (Score:5, Funny)
Grendel wouldn't stand a chance!
Re: (Score:2)
That's 50A at 120V. A very common and standard config is to have two 20A PDUs in a rack, so 50A isn't that much of an increase. You do need an efficient cooling setup, but it's nothing that most commercial datacenters couldn't handle.
Re: (Score:2)
Actually, never mind. I was off by a zero there. That is quite a bit of cooling!
Re: (Score:2, Insightful)
Re: (Score:3, Informative)
By my calculations 1 m^3/sec of air can carry away 65kW at a 50 degK temperature rise. That's doable, though you don't want it exhausting into your office.
A device mostly for C3? (Score:2)
Coupled with the power source this thing on the field would have a huge heat signature that would work like a beacon for enemy missiles I guess, but I will gladly be corrected by any reader that served on the armed forces and have a better idea.
Re: (Score:2)
It's heat signature will be no worse that that of a large internal combustion engine. The enemy already knows the location of division headquarters.
Re: (Score:2)
It would have to be where a network connection of any usable speed is impossible.
BTW, just how big is a steady, 60 KWh solar panel? It appears to require several thousand (around 4,000 100 Watt panels in the mid-atlantic region (and suitable battery infrastructure) to generate 60 KWh for 24 hours a day (1.44 MegaWatt/hours for every 24 hours).
The above is from http://www.batterycountry.com/ShopSite/sec.htm [batterycountry.com]
Re: (Score:2)
It would have to be where a network connection of any usable speed is impossible.
Clearly this supercomputer is intended for an autonomous killing machine... a terminator, if you will. As such it doesn't not need a massive network connection, just enough to control some servos really. At 19" this component would compromise the torso of the machine.
BTW, just how big is a steady, 60 KWh solar panel? It appears to require several thousand
Perhaps, but this machine will no doubt will contain some type of nuclear power core.
Re: (Score:2)
It would have to be where a network connection of any usable speed is impossible.
Having a network connection to a traditional room-filling air-conditioned supercomputer isn't an alternative unless they only want a couple of these, which I doubt is the case (on a side note I'd guess the input bandwidth they need isn't that great - probably just real-time video/voice).
I think it's more a matter of wanting to have this sort of computing power available in the field on a potentially widespread (not one-off) sca
Re: (Score:2)
> ... run Linux?
It probably will.