DARPA Targets Computing's Achilles Heel: Power

coondoggie writes "The power required to increase computing performance, especially in embedded or sensor systems has become a serious constraint and is restricting the potential of future systems. Technologists from the Defense Advanced Research Projects Agency are looking for an ambitious answer to the problem and will next month detail a new program it expects will develop power technologies that could bolster system power output from today's 1 GFLOPS/watt to 75 GFLOPS/watt."

Comments

let me answer that with a question

[FrozenFood]

Do the goverment know of an upcomming energy crysis?

Re:let me answer that with a question

[Anonymous Coward]

No, the problem is getting hold of raw materials for batteries. Mobile computing is on the rise and the west doesn't want to be too dependent on foreign mineral deposits. More efficient computers = smaller batteries = smaller amounts of lithium etc needed.

Re:let me answer that with a question

[unts]

The problem is not just generating the power... it's delivering it and consuming it without breaking/melting. And that's what they're getting at here - getting more FLOPS per watt... not finding out how to push more watts into a system. A silly amount of the energy going into a supercomputer comes out as heat... and a silly amount of energy is then used to remove that heat. Hopefully, by significantly improving the energy efficiency of chips and systems, we can make them a lot more powerful without them needing a whole lot more power. And I haven't even mentioned the mobile/embedded side of the spectrum where its about battery life and comfortable operating temperatures... the same energy efficiency goals apply.

This is the sort of thing we over the pond are very interested in too. Like for example *cough* the Microelectronics Research Group that I'm a part of.

Re:let me answer that with a question

[FooAtWFU]

It occurred to me the other day that, while I have been programming and working with network monitoring tools and the like for a while, and I can get an email alert (or text message) whenever a piece of equipment goes down, the rest of the world doesn't have that sort of capability. A big chunk of of California Highway 1 could fall into the ocean, and people could fall off after it, and no one would notice until someone called it in. If my hard disk is on fire, I can get a message, but if the woods are on fire, you need to wait for someone to see the smoke.

Sensors and the like are pretty awesome to have.

Re:let me answer that with a question

[0123456]

Do you realize how much CPU is required to decode h.264 1080p pr0n? What's the use of a laptop without it?

We used to that on a 200MHz dual-core ARM with some hardware decoding assist. If I remember correctly the whole system used less than 1.5W and much of that was the video encoder for the TV (we were using analogue component at the time, not HDMI).

And with GPU assist an Atom with a low-end GPU can happily play 1080P H.264.

Re:let me answer that with a question

[stevelinton]

In a sense. There is a widespread view that we will need 1 Exaflop supercomputers by roughly 2019 or 2020 for a whole range of applications including aircraft design, biochemistry to processing data from new instruments like the square kilometer array. On current trends, such a computer will need gigawatts of power (literally), which amongst other things would force it to be located right next to a large power station that wasn't needed for other purposes. This is felt to be a bit of a problem and this DARPA initiative is just one small part of the effort to tackle this and get the Exaflop machine down to 50MW or so, which is the most that can be routinely supplied by standard infrastructure.

Re:let me answer that with a question

[Teun]

Why do you need an energy crisis to make something work more efficiently?
Concidering energy does not come cheap there is a very good commercial reason to save on one of the larger costs in computing (or any other activity)

And even though the US hosts the worldleaders in denial of CO2 related climate change it is still an ever more important concideration for many people, even in the US.

Re:let me answer that with a question

[Luckyo]

Problem with lithium, it isn't mushroom and berries. You can just walk in there and pick it up. It's also not oil. You can't just put a hole in the ground, connect it to the pumping machinery and have oil. You need to have an actual ore mines, with huge, easy to sabotage, hard to fix machinery.

And finally, it's solid and heavy. It's a total bitch to move from center of war-torn nation that has world's best specialists in asymmetric warfare fighting against you both economically and in terms of general feasibility.

Re:let me answer that with a question

[Anonymous Coward]

In a pinch you can extract lithium from sea water. That's basically what a lithium deposit is... an old sea that dried up and left the salts. Lithium isn't a big fraction of a battery's cost, weight or volume. Please everyone stop being silly. The cobalt that is often used in lithium batteries is far more expensive, rare and used in larger proportions. We just don't call them cobalt batteries so no one knows about that part.

Re:let me answer that with a question

[Rockoon]

With an exaflop computer, simulating the human brain is looking like it might be possible.

Take a moment, relax, and then try to answer this question: What does computational speed have to do with it?

The point is that simulations are not linked to computational speed. Some simulations that we do today are performed thousands of times faster than "reality" while most others that we do today are performed thousands, or even millions of times slower. The speed of the simulation is irrelevant to their existence, so stop pretending that speed has any sort of importance to simulating something like a brain. A turing machine is a turing machine. Period.

Pretty much every day of your recent life you have observe the results of simulations that are trillions of times slower than reality, and that is the simulation of the propagation of light. We dont need faster computers to simulate a brain.. we just need to figure out how to simulate a brain.

tl;dr: The parent shouldn't talk about things that he doesnt understand.

Re:let me answer that with a question

[SeaFox]

And with GPU assist an Atom with a low-end GPU can happily play 1080P H.264.

Actually that depends on the bitrate of the encoding far more than whether it's "1080p" or not. I've seen plenty of "1080p H.264 video" that's got lousy quality the moment there's any action.

Not to mention the what profile of h264 was being used. High Profile requires much more computational power than Main. We're also assuming the video can be GPU accelerated. You can't just take any h264 video and get hardware acceleration, the video has to be encoded following certain rules about bitrate, b-frames, etc otherwise it will be all decoded in software.

Re:let me answer that with a question

[ultramk]

Erm, not to be overly pedantic, isn't *all* of the energy consumed by a supercomputer (or any other device) eventually converted into heat? First law of thermodynamics and all that?

Turing Tax

[Wierdy1024]

The amount of computation done per unit energy, isn't really the issue. Instead the problem is the amount of _USEFUL_ computation done per unit energy.

The majority of power in a modern system goes into moving data around, and other tasks which are not the actual desired computation. Examples of this are incrementing the program counter, figuring out instruction dependancies, and moving data between levels of caches. The actual computation of the data is tiny in comparison.

Why do we do this then? Most of the power goes to what is informally called the "Turing Tax" - the extra things required to allow a given processor to be general purpose - ie. to compute anything. A single purpose piece of hardware can only do one thing, but is vastly more efficient, because all the power used figuring out which bits of data need to go where can all be left out. Consider it like the difference between a road network that lets you go anywhere and a road with no junctions in a straight line between your house and your work. One is general purpose (you can go anywhere), the other is only good for one thing, but much quicker and more efficient.

To get nearer our goal, computers are getting components that are less flexible. Less flexibility means less Turing Tax. For example video encoder cores can do massive amounts of computation, yet they can only encode video - nothing else. For comparison, an HD video camera can record 1080p video in real time with only a couple of Watts. A PC (without hardware encoder) would take 15 mins or so to encode each minute of HD video, using far more power along the way.

The future of low power computing is to find clever ways of making special purpose hardware to do the most computationally heavy stuff such that the power hungry general purpose processors have less stuff left to do.

Re:Turing Tax

[Anonymous Coward]

For comparison, an HD video camera can record 1080p video in real time with only a couple of Watts. A PC (without hardware encoder) would take 15 mins or so to encode each minute of HD video, using far more power along the way.

While it makes your point, you're actually off by orders of magnitude on both: a modern PC can easily encode at 2-4x realtime for 1080p... and a good hardware encoder often uses less than 100 milliwatts. A typical rule of thumb is that dedicated hardware is roughly 1000 times more efficient, power-wise, than a CPU performing the same task.

Re:Turing Tax

[Kjella]

To get nearer our goal, computers are getting components that are less flexible.

Actually, computers have lost lots of dedicated processing units because it just wasn't worth doing in dedicated hardware, that's where for example softmodems (aka winmodems) came from. And with GPUs going from fixed pipelines to programmable shader units, they too have gone the other way. Dedicated hardware only works if you are doing a large number of exactly defined calculations from a well established standard, like say AES or H.264. Even in a supercomputer the job just isn't static enough, if the researchers have to tweak the algorithm are you going to do build a new computer? You have parameters, but the moment they say "oh and we have to add a new correction factor here" you're totally screwed. Not going to happen.

they should talk to TI

[Gravis Zero]

TI's line of MSP430 chips run using little solar cells. hell, they practically run on their own self-esteem. so scale that technology and bam, you got a super computer that runs on a couple AA batteries.

But you loose flexibility

[SmallFurryCreature]

If you want to talk about encoding, anime fan subbers are at the fore front. The latest is 10 bit encoding. It has a lot of benefits but what its main downside is that there is no hardware for it, you need to run it on the cpu. Someday hardware like a GPU might support it but that takes far to long to stay current.

That is the reason the general purpose CPU has won out so far, why mobile phones and tablet come with them as the main computing unit, because keeping up in hardware with the latest developments just is to slow.

You could in theory build a super computer that can run ONE task very fast. They existed, in fact the earliest computers WERE single task machines... and they lost out because the next task might be totally different and building a new machine for each task is slow and expensive.

The person below (Wierdy) talks about one bit of modern codecs... but this might change tomorrow, as indeed it has with 10bit encoding.

There is a reason DVD's suck donkey balls, open one up and look at what is inside and wonder why the fuck any of it was needed when any PC could easily have dealt with a better format (files max size, subtitle format etc)... because DVD players were purpose build devices and had to be designed ahead of current techonolgy to be widely supported. DVD players being purpose build single task hardware started out obsolete and couldn't change. Of course, the advantage was the they were relatively cheap and became cheaper BUT do you REALLY want your super computing to be this inflexible?

In many ways, the current GPU craze is nothing more then math co-processor of yesterday, or the windows chip on early video cards. They are useful but can't stay up with the rapid advances software can make.

The real money is in making generic hardware faster and more efficient because that is where the intresting stuff is happening. Profit wise as well. What would you rather be selling, DVD players or iPads?

Power Consumption

[AnotherAnonymousUser]

Is there any sort of rule of thumb when measuring power consumption - ie, X amount of processing uses Y blocks of power? Is there a theoretical minimum requirement of energy to perform certain types of calculations?

Re:Power Consumption

[alreaud]

Yes and actually very simple (SI units),

P = C*V^2*f where P is power in Watts, C is capacitance in farads, V is voltage in volts, and f is frequency in Hertz. C is kind of hard to measure, and is dynamic depending on processor load. A design value can be determined from processor data sheets.

Power is only consumed in MOS transistors during transitions, to the value I = C*dv/dt, where C is the overall transistor capacitance to the power supply, in this instance. If dv is 0, ie, at a stable logic level, then I must also be 0, and hence power dissipation must be zero due to Ohm's Law, P=I*V.

At 1.3V and 2.8GHz, the dv/dt multiplier becomes 4.73*10^9, significant for a 100-million transistor microprocessor even if overall capacitance/transistor ~femtofarads.

Re:Computronium

[Curunir_wolf]

Getting the level of detail from a brain you'd need to simulate it might be less a matter of implants than destructive readout. Slice-and-scan.

Well, right. That also eliminates the potential issues from having duplicate persons in virtual space and meat space.