back to article IBM's 'neurosynaptic chip' to power nuke-watching exascale rig

IBM's TrueNorth platform will form the basis of a collaboration with Lawrence Livermore National Laboratory (LLNL) to chase the exascale dream. The “neurosynaptic chip” Big Blue – or perhaps, given its much-hyped “brain-inspired” design, we should say “Big Grey”? – will “process the equivalent of 16 million neurons and 4 …

  1. Conundrum1885 Bronze badge

    So

    This is very similar to the way the human brain works, ie all the circuits are "there" but only become active when needed.

    How have they overcome electron resistance across the neural filaments? Silver interconnects?!

    I've got some ideas here for a silver/iridium alloy which might work, at least as far as preventing the troublesome formation of conductive silver oxide.

    1. Anonymous Coward
      Anonymous Coward

      Re: So

      Well... obviously :)

  2. allthecoolshortnamesweretaken

    How does the chip compare to a human brain in terms of power consumption?

    1. Alan.Griffiths.2013

      Re: relative brain power

      1 million neurons == a cockroach == 0.00001163 human

      1. raving angry loony

        Re: relative brain power

        So if it's got the brain power of a cockroach does that mean it'll either outwit many humans or become a (real) estate agent?

    2. PleebSmasher

      16 TrueNorth = 2.5 W. Human brain = 20 W. Assuming constant overhead, you could use 128 TrueNorth and 20 W.

      The power consumption of this chip is so low that I would expect to see 3D/stacked versions of the chip in planning, something Big Blue has always been interested in.

      1. allthecoolshortnamesweretaken
        Pint

        Thanks, have one on me!

      2. Anonymous Coward
        Anonymous Coward

        Yeah, it is kind of embarrassing how little we are able to do with all that horsepower.

  3. Anonymous Coward
    Anonymous Coward

    Skynet

    It's OK, the article doesn't mention Miles Dyson, or Cyberdyne systems anywhere.

    1. Crisp Silver badge
      Gimp

      Re: Skynet

      I was wondering what those chips reminded me of...

      It almost looks like a component from a Cyberdyne Systems model 101.

      1. Johndoe888
    2. Anonymous Coward
      Anonymous Coward

      Re: Skynet

      That's becasue they are not allowed to.

      It's a conspiracy, man!

    3. allthecoolshortnamesweretaken
      Coat

      Re: Skynet

      Dyson? What do vacuum cleaners have to do with this?

  4. Pascal Monett Silver badge

    2.5 watts ? Really ?

    Is there a kilo missing there ? For 60k+ cores, it wouldn't be unreasonable to use 2.5 kilowatts during calculations - at least it would seem very reasonable to me for that number of chips.

    The most expensive Intel chip (Xeon E5-2670 v3 I think) is rated at 120W thermal dissipation. It has 12 cores. That makes 10W per core. Scaled up to 64000 cores, that is 640 kilowatts.

    So are we really talking about a mere 2.5 watts ? Not even enough to boil water, I'd wager. If that is the case, then count my mind blown. And even at 2.5 kilowatts, it would still be mightily impressive.

    1. Crisp Silver badge

      Re: 2.5 watts ? Really ?

      Apparently both the cores and the on-chip network operate without clocking.

      That, IBM says, is the key to the low power consumption in TrueNorth.

    2. PleebSmasher

      Re: 2.5 watts ? Really ?

      There's no mistake. It really uses just 2.5 Watts or 70 mW per individual TrueNorth chip.

      Neuromorphic computing is supposed to be brain-inspired. It's supposed to simulate spiking neurons with a massively parallel architecture. The brain uses roughly 20 W to accomplish this.

      The first challenge is in programming the thing or proving that it is useful at all, Apparently some think it is useful. I would be shocked if Google/Microsoft weren't laying down a few millions to adapt machine learning to neuromorphic chip designs. Scaling it up is also a challenge. If TrueNorth was cheaper, could you buy the equivalent of 100 billion "neurons" within a reasonable budget? When will a version of TrueNorth use 3D stacked transistors/neurons to increase the density to something comparable to a human brain? V-NAND already has 48 layers and more coming. A 16-chip TrueNorth board with 1 billion "neurons" could be within IBM's reach.

      1. JeffyPoooh Silver badge
        Pint

        Re: 2.5 watts ? Really ?

        PS "The first challenge is in programming the thing..."

        It's a neural net. Most of us have some experience programming such things...

        It'll take about 20 years. Or 25 years if it happens to adopt the Goth lifestyle along the way.

        Hopefully it'll keep its room tidy and not back talk too much.

    3. SKC

      Re: 2.5 watts ? Really ?

      http://www.research.ibm.com/articles/brain-chip.shtml

      'The chip consumes merely 70 milliwatts'

      'Organic wetware' wins the phrase of the day too. Proper cyberpunk methinks.

    4. Troy Peterson

      Re: 2.5 watts ? Really ?

      These are not x86 derived cores. Consider that they are designed specifically for the task at hand and probably only have a small handful of instructions... If they are just emulating neurons then op-amps wired as comparators with configurable biases are a close approximation to a neuron... However I suspect that these are digital cores which only have to run a very small 'program' at extreme parallelism like the shader units in a GPU... So if you consider a processor core that is dead simple without all of the fancy pipelining and out-of-order execution of modern CPUs that has only a handful of instructions and cascade-switches instead of being clocked... then 70mw is very reasonable...

      1. Dave 126 Silver badge

        Re: 2.5 watts ? Really ?

        >So are we really talking about a mere 2.5 watts ? Not even enough to boil water, I'd wager.

        How much water?

  5. inmypjs Silver badge

    2.5W what does it mean?

    Event driven clockless logic? The power consumption (after leakage) then depends on the rate events occur. This and the linked article make no mention of event rates or event propagation speeds.

    They brag about thousands of cores and billions of transistors but the 2.5W really means most of those cores and transistors are not doing anything most of the time.

    It will be interesting to see if they can come up with brain like solutions to useful problems and how performance compares with traditional architectures.

    1. Aremmes
      Holmes

      Re: 2.5W what does it mean?

      "They brag about thousands of cores and billions of transistors but the 2.5W really means most of those cores and transistors are not doing anything most of the time."

      Just like how we only use 10% of our brain at any given time.

      1. Mike 16 Silver badge

        Re: 2.5W what does it mean?

        -- Just like how we only use 10% of our brain at any given time. --

        Sort of, if that were true. (Anything follows from falsehood).

        Meanwhile, for those thinking "Xeon", perhaps you might consider "Connection Machine". After all, it did a impressive job of allowing dinosaur reanimation, while ignoring high-school level biology about amphibian gender-bending. :-)

      2. Anonymous Coward
        Anonymous Coward

        Re: 2.5W what does it mean?

        If we "used" 100% of the brain, it wouldn't work.

        So the 10% brain usage is a silly old myth. I thought everyone knew that by now.

  6. NanoMeter

    Skynet will not arrive

    until version 5, with 100 times the power. Then we're in bloody trouble.

    1. D 13
      Terminator

      Re: Skynet will not arrive

      A hundred times is only going to get you to a Golden Hamster, unlikely to present too much of a threat to the future of humanity.

      At 10^13 you get cats, which could go either way.

      100 times more and you get humans, which will probably present a challenge.

      Just four times more than that and it would be smarter than an African Elephant, then we're fucked.

      https://en.wikipedia.org/wiki/List_of_animals_by_number_of_neurons

      Disclaimer - this is just what I looked up in Wikipedia while reading the comments, not a clue if it's accurate or not.

      1. PleebSmasher

        Re: Skynet will not arrive

        Larger animals have more neurons needed for movement.

      2. Crisp Silver badge

        Re: Skynet will not arrive

        Thanks to that wikipedia link, and checking brain to body mass ratios I found this:

        The bony-eared assfish.

  7. Dave 126 Silver badge

    Informative comment from another site:

    bluemellophone

    3/30/16 11:49am

    Alright, Ph.D. computer vision researcher here.

    First off, TrueNorth is a great project. Our university has been able to get its hands on a few of these chips for testing - the folks working on these chips are about a 10 second walk down the hall from my lab. TrueNorth has turned the now-somewhat-routine computer vision research problem of image classification on its side by approaching it from a different angle: the hardware. This is great because nobody else is really doing this on a large scale except for IBM. TrueNorth could lead to some neat new insights on how to make our current solutions more computationally and memory efficient. In some aspects, it already has. That’s not to say that TrueNorth is limited to only computer vision applications, but it is why I’m curious about its recent developments.

    That being said, TrueNorth has by no means the same level of reliability, accuracy, or scalability of the technologies behind Google’s self-driving cars or Facebook’s face detection or Microsoft’s Xbox Kinect. The latest research (http://arxiv.org/abs/1603.08270, http://arxiv.org/abs/1602.02830) indicates that TrueNorth has a difficult time implementing a particular operation called a convolution. Convolutions are important because it allows for a computer to take a large, complex image — of say, a cat — and boil it down to its most important conceptual components — like fur, cat ears, tail. There is evidence that our brains work in a similar way to deconstruct an image into its abstract concepts so that our brains can process what we see. This is a problem for TrueNorth because the cool, sexy computer vision applications making the recent headlines are pretty much all based on Convolutional Neural Networks (CNNs). Specifically, TrueNorth implements a form of CNN known as a BinaryNet by Courbariaux et al. but with some pretty severe technical drawbacks.

    Long story short, TrueNorth may someday make its way onto phones for select tasks, but take the GIF at the top of the story with a grain of salt. The development of this platform is in its infancy. Another platform to watch is Nvidia’s Jetson line, which has an architecture more akin with ongoing research in the field and thus can inherit state-of-the-art ideas easier. I’m interested to see where TrueNorth ends up in 5 years, but I’m not holding my breath for the field to adopt it en masse.

    - http://gizmodo.com/this-supercomputer-mimics-a-human-brain-using-just-2-5-1767958119

  8. JeffyPoooh Silver badge
    Pint

    "...16 million neurons and 4 billion synapses just 2.5 watts..."

    If you log the state and process at 0 Hz, then you effectively put the drive on a shelf and walk away....zero power.

    You could do the same thing with 16 billion neurons.

    Or a 16 trillion.

    1. Conundrum1885 Bronze badge

      Re: "...16 million neurons and 4 billion synapses just 2.5 watts..."

      Its intriguing to note that on my desk I have a graphics card which 4 years ago was state of the art, containing at the time 2GB RAM and being about twice as fast as a PS2's EE CPU, costing £110.

      That same specification was as of November 2015 being given away free with a magazine, draws 1/10 the power and actually runs faster than this card with 1.6* the raw IOPS.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2019