Slashdot Mirror
Stanford Bioengineers Develop 'Neurocore' Chips 9,000 Times Faster Than a PC
kelk1 sends this article from the Stanford News Service:
"Stanford bioengineers have developed faster, more energy-efficient microchips based on the human brain – 9,000 times faster and using significantly less power than a typical PC (abstract). Kwabena Boahen and his team have developed Neurogrid, a circuit board consisting of 16 custom-designed 'Neurocore' chips. Together these 16 chips can simulate 1 million neurons and billions of synaptic connections. The team designed these chips with power efficiency in mind. Their strategy was to enable certain synapses to share hardware circuits. ... But much work lies ahead. Each of the current million-neuron Neurogrid circuit boards cost about $40,000. (...) Neurogrid is based on 16 Neurocores, each of which supports 65,536 neurons. Those chips were made using 15-year-old fabrication technologies. By switching to modern manufacturing processes and fabricating the chips in large volumes, he could cut a Neurocore's cost 100-fold – suggesting a million-neuron board for $400 a copy."
Are you ready?
If they can use modern fabs, then we will have a simulate brain in a decade.
The Kruger Dunning explains most post on
Come back with something newsworthy.
Misleading headline.... these obviously won't be 9,000 times faster than PCs at most things. They do one thing only, and that's simulate a neural network.... Slashdot is now like a tabloid...
I've just seen three comments deleted within 5 minutes complaining about the lack of journalism.... Copy and pasting a press release headline without any real reporting is what tabloids are for.
but will it play Crysis?
that said, what sort of memory (short-term storage) is used? Would that be the current bottleneck?
It would definitely be interesting to see how this continues to develop.
It might be too late for bitcoins but perhaps one of the altcoins can benefit.
Or for weather prediction/modeling.
Then again, the dark side comes to mind to (skynet, SID 6.7, etc.)
Soooo, it's a design "based on the human brain" that then "simulates a human brain". Meaning its an ASIC like chip so of course its faster than a general CPU.
not to forget, the female brain one will crash and be near unusable for several days every month.
Over 9000 times ?? That cannot be just a co-incidence
At last: a computer that will be as frustrated by computers as I am!
Your thinking of the vagina. Dont feel bad, We all think of the vagina all the time.
or vagina. Sorry, what were you saying again?
Is my CPU going to struggle with depression and anxiety now?
Good old clueless tech journalists, followed by slashdot editors just copy pasting.
The chips aren't 9000 times faster than a typical PC for general tasks. Specifically, they can simulate neurons 9000 times faster than a PC can simulate neurons. Pretty typical of any ASIC with a limited set of a highly specialised functions.
Another subgenre of EDM.
Actually, randomly connecting neurons can be functionally useful. Some (somewhat crazy) theorists think that random projection is an underlying principle of the brain; a mathematical concept that is useful for dimensionality reduction leveraging special crafting of random matrices, so they aren't completely random.
Cray Cray.
Never answer an anonymous letter. - Yogi Berra
I doubt it. Well, at least not as soon as you might imagine. "Together these 16 chips can simulate 1 million neurons and billions of synaptic connections"
Total number of neurons in cerebral cortex =
--10 billion (from G.M. Shepherd, The Synaptic Organization of the Brain, 1998, p. 6).
--20 billion (Biophysics of Computation. Information Processing in Single Neurons, New York: Oxford Univ. Press, 1999, page 87).
Total number of synapses in cerebral cortex
-- 60 trillion (from G.M. Shepherd, The Synaptic Organization of the Brain, 1998, p. 6).
--150 trillion (Pakkenberg et al., 1997; 2003)
--240 trillion (Biophysics of Computation. Information Processing in Single Neurons, New York: Oxford Univ. Press, 1999, page 87).
So, lets call it 15 billion neurons and 150 trillion synapses, or tens of thousands of synapses per neuron, ten times as many as this chip provides. That's going to be a problem. To say nothing of the fact that I would be very surprised if it allows for billions of inter-chip synapses which would probably be necessary to model the non-local interconnections common in the brain within the 240,000 chip brain simulator. And that's just for the cerebral cortex. You've got the rest of the brain to simulate as well.
Then there's the glial cells, which outnumber neurons by 10-50:1, and which recent research suggests may be considerably more involved in neural activity than presumed by the traditional "life support and other infrastructure" understanding.
Could be great for modeling larger portions of a mouse brain though. Maybe even to start modeling the simpler parts of a human brain. And we do have to start somewhere. I suspect we're at least a few decades away from being able to begin to simulate an entire human brain, and probably many more decades away from getting the simulation accurate enough that it might begin to actually function properly. After all the number one benefit of these simulations is to fail spectacularly in interesting way in order to help neuroscientists figure out what questions they should be asking.
Meanwhile we need to ask ourselves - if we're creating this simulation based on the human brain, then what are the odds that some form of consciousness dwells within it? And what sort of torture are we subjecting it to as it's simulation collapses? And does the knowledge we gain justify that price?
--- Most topics have many sides worth arguing, allow me to take one opposite you.
Not that I'm knocking it. A GPU implements specific algorithms to great effect. But a GPU's algorithms are ones that are interesting for a specific application (drawing texture-mapped polygons), whereas an artificial neural network still needs another layer of programming to do something useful. In other words, a Word Processor implemented on this chip would not be 9000x faster than a Word Processor implemented on a CPU. A face recognition algorithm, on the other hand, might see a decent fraction of that 9000x, although it remains to be seen whether this chip would be a better fit for any particular application than a GPU (for example).
It isn't a typical ASIC; the chip is a custom fully asynchronous mixed digital+analog; the board uses 16 chips in a tree router for guaranteed deadlock prevention between the chips; and can simulate 1 million neurons powered only by one USB port.
The neurons are implemented with analog circuits to match the dynamics of real neurons, moving beyond a simple hodgkin-huxley model to include components like ion channels, which is first of its kind in an analog chip. It has a neat hexahedral resistor network that distributes the spike impulse across a neighborhood of neurons, a phenomena seen in many cortical brain areas; essentially an analog phenomena implemented efficiently in analog design.
Analog gives it fun biological-like properties, with things like temperature sensitivity that must be regulated with additional circuitry. Asynchronous design means outside of leakage from the chip, which is low with such a large fabrication process, very little energy is used at a neuron level if no stimuli is present. This is in contrast to a traditional CPU, which has a clock marching along lots of a chip to consume energy every clock cycle.
Outside of wireless/signaling stuff, this is probably the biggest mixed analog digital asynchronous chip in existence.
But otherwise yes, the editors sucked on this one.
All models are simplified. This simulator happens to incorporate ion channels, and other effects, and has been used to replicate many real world behaviors.
I would have to do more research but I thought we already had done some (not in real time... but close?) simulations of sections of the human brain already. The only parts I remember from those articles was that it was for researching the visual processing for military applications I think (Luke binoculars I think?). I also remember that this project had done a seemingly realistic simulation of a mouse and cat brain in total. From what I understand the simulations supposedly acted as a "real" one would. Sorry for being vague. I am a layman and that is just what I remember.
GTA IV and Kerbal Space program with no lag!
Donald Trump, on a crusade to make Nixon look respectable
Misogyny. Even as a man I find this to sometimes be overly repetitive.
in this story that was self-published to Kindle in 2010:
http://www.amazon.com/America-The-Enslaved-Neurochip-ebook/dp/B007LAX6YY
It’s very very different; nerumorphic chips have been around for ages, they use the same phenomena the brain does (ion-flow across a neuron's membrane) using different a method (electron flow across a silicon membrane).
The big difference is that they make use of analogue computation using the physical properties of electricity to model whatever you’re trying to model, whereas digital computers model things by representing quantities as symbolic values.
So digital computers let you model something by simulating it with symbolic values, an analogue computer lets you model something by emulating a systems physical properties.
There is no machine code to speak of, you can’t program an analogue circuit, you have to physically construct it. That’s what makes this Neurogrid technology is interesting; if these guys are on the level then they've developed a practical way to use digital computers to “program” analogue circuits.
That looks nice, but what problem does it solves?
Isn't this somewhat (at least loosely) similar to the approach that Thinking Machines took with the Connection Machine CM-5? I realize it's very different, but the rationale is the same. Granted, they're putting on a single die what the CM-5 was in totality.
they could have said it was over 9000 faster than a PC
I have read so many great things here over the years slashdot has existed but a link where you can get it never ever appears.
By the time this is ready a new chip we will certainly be reading about.
In 1989 I was doing billions of connections per second on DataCube finite impulse response filter hardware to do the weighted sums, and hardware look up table for the sigmoid mapping for trainable multisource image segmentation for around $40,000 in off-the-shelf VME bus hardware, but that was in 1989 dollars, so I guess there has been some advancement.
Seastead this.
1 million neurons. ~1000 inputs each.
Alot of hype because maybe those large sounding raw (limited) processing operation counts (and not floating point math of any accuracy) are Apples to the normal PC's (a general purpose processor) Oranges.
The reader might do a small sampling of NN Neural Net programming difficulties, where its been shown that there are limited problem sets that all neural nets are applicable to (and this IS only one limited flavor) and then the problem of forming the thing's "program" (the real difficult part) where the likelihood of failure/trashing grows exponentially with the size of the 'net' (complexity of the problem it is meant to solve).
Training any neural net is its chokepoint and EVERY problem domain has to be tediously trained again (alot of human intercession) - and thats for problem sets that you CAN provide ALOT of training data for (and not large sets of unknowns).
If you remove the outliers, our brains seem to be growing in accordance with Moore's Law.
I wonder how many there are now...
How fast can it get me bitcoins? OMGZ I need to buy one now. Wait. After reading the article, it's pretty much ambiguous nonsense.
Human brain is 100 billion neurons. $40 million for human equivalence.
And we still don't know how to download Johnny Depp.
I'd love to see a beowulf cluster of those!
This might be yet another step on the way to a truly sentient artificial intelligence - but even if these things live up to their promise, we still have a long way to go before we artificially create consciousness. Incidentally, how will we know that we've succeeded? The old "ability to ask the question is the answer" rule doesn't apply, as the device could easily be programmed to ask - or might just randomly ask, but not really care about the answer.
Creatures with fewer neurons typically have more specialization per neuron, with fewer neurons that are similar. In a weird way, having many homogeneous neurons may make understanding the brain easier.
ASIC's are faster than software!
Stop the press everybody!
It's good you like that, because they absolutely are treated differently in the industry and academia when you are working on circuits that are fabricated in CMOS.
The process of designing and testing a digital system is much simpler than one where you need to incorporate analog dynamics into your model. With a digital designs, teams of engineers for the fabrication companies (like TSMC or Intel or Global Foundries) and CAD companies (Cadence, Synopsis) prove that your signals are stable within the right timescales for your digital system, and so all you need to do is build on top of that without questioning the analog dynamics underneath. With analog, this isnt true. In fact, many cases you need to do special tricks to have your analog circuits fabricate properly because you arent using one of the standard components, which breaks assumptions that the fabrication companies may have. Analog systems have continuous range of inputs, so the testing space is even larger than digital. But I'm sure you already knew all about that :)
IT'S OVER 9000!!
Neurons have incredibly complex behaviors, they are not simply threshold triggers as the simple CS model implies.
You're plainly ignorant. I don't have any threshold triggers in any of my neural networks. Cells have complex protein behaviors, so what? The cybernetic models can be Turing complete. This means that if I really wanted to waste CPU power instead of understanding the fundamental principals of cognition, I could build a neural network that emulated the molecular action of cellular proteins, and if our rate of computer advancements holds that machine intelligence would be able to emulate the molecules that make up human neuron proteins, and eventually an entire human head right down to the molecular level. Artificial neural networks can yield every bit as much complexity as anything else in nature. Did you forget that electrons are made of quantum particles or something? Now, we're shooting for determinism and thus applying quantifications in most cases, but in the future we'll harness things like eddy currents once our n.net model methodologies have nailed down and abstracted more of the key components that emerge of complex behaviors efficiently.
Neural networks in CS have little to do with the actual wiring and primarily chemical systems that are neurons.
Nor do the artificial neurons need to have anything to do with organic ones except very basic fundamental properties which produce the complexity of response and thus intelligence. I suppose next you'll be telling me that without putting a human brain in the boxen we won't be able to make personal computers do mathematics.
You are what I call an organic chauvinist. What's so damn special about the precise chemical functionality of organic brain operations? If the organic chemputers were such a grand and complex design in need of exact duplication to achieve any degree of similar intelligence, then why are dumb computing machines even able to revolutionize computation? How are digital cameras doing facial recognition with far less computation power than human brains require? It's true that organic neurons have more internal state and some of the details of the process by which neurons operate are still undiscovered; However, we don't need to achieve the exact nuanced behavior of human neurons or even the same human brain neuron capacity scale or even its same connectivity types in order to produce intelligent behaviors. There are some general principals at work that any complex system will exhibit in order to achieve a given behavior, and those are worth emulating in an optimized fashion. Nature has converged upon solutions randomly using trial and error and going with the first working attempt the entropy gives her whether it is optimal or not. Replicating every detail of said accidental functionality exactly is not essential any more than it is essential for creatures to have 4 legs in order to walk.
It's already been proven that complexity yields intelligence. The more neurons the smarter the entity. In fact, we have been determining the minimal degree of complexity required to solve various problems, and nearly universally we can solve the same problems with far less complexity than the equivalent solution in nature, since organisms weren't intelligently designed. There is no binary dichotomy: An interaction does not reach some threshold and then magically becomes intelligent. Instead, there is an intelligence gradient: All systems exhibit some degree of "intelligence" AKA processing power, and the amount scales with complexity. Even a run of dominoes has some small degree of intelligence. Human brains have a lot of neurons doing stuff that isn't even required to produce sentience (thermal regulation, breath control, motor skills, etc). In fact, you can take whatever estimate your cognitive neuroscience prof claims the human brain has as a yardstick for the complexity requirement of sentience and
That was Mentifex quality.
At 400$ a pop, I'd be willing to shell the cash to have access to this kind of chip/board. There's at least one direct application I'd like to try: source code analysis. The current tools are quite powerful, mind you, but I'm sure the pattern recognition capabilities of such chips should be a lot better at pinpointing ill side effects, inefficiencies, memory leaks and such.
Now, just imagine a biowolf cluster of those...
>You'd sooner have "cured" the CERN researchers of "Particle Fever" [youtube.com], and the attempt to dissuade interested CS folk from becoming cyberneticians would be equally as foolish as decommissioning the Large Hadron Collider before discovering the Higgs boson.
You completely misconstrued my statement and argued to a conclusion above that resembles nothing of the original precept. I believe in machine emulated neural computation. For all your rambling, you failed to get that the neurons embodied on these ASIC chips are NOT the kind needed to perform brain-like computation, and yes, including the complex connections of axons and dendrites. If anyone is going to produce simulations of cognition, it will be CS folk. The problem is that non-cognitive educated CS folk see primitive neuron (ie. SINGLE neuron) models as definitive models of organic neurons. They are NOT, and you know that. Different neurotransmitters change the nature of function around neurons - similar functionalities must all be simulated in some form, but you must first acknowledge they exist, something the basic CS crowd have yet to figure out.
>So, you're going to have a VERY hard time "curing" me of the machine intelligence bug
I'm not trying at all - quite the opposite. Nobody talked about curing you of anything. You simply wasted your time misunderstanding the original post. I believe sternly in machine transhumanism and it is only a matter of time. Now get back to work on the problem and stop wasting everybody's time.
Stanford Bioengineers Develop 'Neurocore' Chips 9,000 Times Faster Than a PC
First, as everyone has already pointed out, they won't be just plain "faster than a PC." They are custom chips designed to do a specific job, so it's not that surprising - if it's true.
Because the article's lead-in reads:
Stanford bioengineers have developed faster, more energy-efficient microchips based on the human brain – 9,000 times faster and using significantly less power than a typical PC.
which is a little unclear as to whether the latter half refers to the new chips or to the human brain.
And then the article says:
The modest cortex of the mouse, for instance, operates 9,000 times faster than a personal computer simulation of its functions.
which again does not refer to the new chips.
systemd is Roko's Basilisk.
"Good old clueless tech journalists" ,followed by slashdot editors just copy pasting"
Profit?
Is that you, Miles?
What's so damn special about the precise chemical functionality of organic brain operations?
Look, until you manage to understand consciousness precisely, there's lots of room for more complexity there than what you normally think of when you talk about a chemical reaction. For instance, we keep finding out that our senses are based on quantum effects. What if it turns out that consciousness is dependent on them as well? At minimum you'd need a quantum computer to get those results. And doesn't intelligence as we understand it require consciousness? Do we have some massive parallel iterator, or is there something special about consciousness that enables intuitive leaps that computers without it may never be able to make?
Luckily for us meatbags, we still don't know.
"You're right," Fisheye says. "I should have set it on 'whip' or 'chop.'"
I wonder if this kind of mass parallel computing architecture, can lead to more efficient ray tracing algorithms.
ya ya, here come the 'quantum' god-heads,. pull the other one, bro
AC: It's not an invocation of "quantum god-head" to state the fact that quantum behaviors are observed in our sensory and perception organs and that we probably need a better conception of quantum mechanics to match some of the computation aspects of human beings.
Dinkypoo: We don't know if there is anything special about the brain and its particular computation structure, but we're making progress on a lot of fronts very rapidly. I think the summary of the long post is that *thus far* nothing about the brain chemistry has stood out as fundamentally unsupportable by silicon and other forms of computation. And even if we have to maintain quantum states to achieve sentient machines, that doesn't mean that we necessarily will have to do it in the same way that the brain chemistry does. I think that's the main thesis of the long post and that it holds, even when considering the observed quantum effects.
"1 million neurons powered only by one USB port"
Yes, but who powers their neurons by USB these days. Most drouds aren't compatible with the USB spec, so simulating USB-powered neurons seems a bit silly.
...who thought WOO FINALLY NO LAGGIES when playing MMORPGs?
Sounds like somebody's looking for some grant money.
You are making good points here -- but nobody was arguing them on this thread.
I'm glad you are really good at cybernetician. It seems like you've been waiting a LONG time to pounce on someone stepping on your turf.
the attempt to dissuade interested CS folk from becoming cyberneticians would be equally as foolish as decommissioning the Large Hadron Collider before discovering the Higgs boson.
I'll bet anyone a dollar that the Large Hadron Collider will not discover the Higgs boson. And anyone dissuaded from being a cybernetician based on someone saying; "Neurons are not simple threshold computation devices" is going to be dissuaded by other things on a regular basis.
>>"ad space available -- low rates!!!"
To bolster the appreciation of organic processes -- I'll say that one cell in the human body has more capabilities and complexity than any single factory on the planet yet created by man.
Grab a few million base pairs while folding and copying the blueprints, construct any one of a million organic molecules, repair itself, and then requisition more materials all on the head of a pin with room for a few thousand more factories? Oh, and while protecting itself from countless biological saboteurs and access attempts by nature trying to disrupt it's processes. It's like having a factory tour where every guest could be strapped with a bomb -- talk about distractions.
>>"ad space available -- low rates!!!"
I agree with you about the quantum processes. It's kind of like attaching "nano" to anything small including chemistry. Quantum processes might be part of every-day and ordinary events like using a compass to find the magnetic north pole -- in fact, Quantum has to be part of nature because everything is built on it by necessity.
And there are many organic processes that while complex, are not efficient. Wires can transmit signals many times faster than our nervous system for instance.
I'm in no hurry for a Sentient Machine however -- we've already got Database technology and mass communications being abused by an Oligarchy to preserve their status quo. Every military advancement I now witness, I worry more about it being used on me -- not defending my precious Citizen status.
Until we can create a just society where some Billionaire doesn't have MORE JUSTICE than I do -- I'm not sure I want Artificial Intelligence in the hands of psychopaths. So while I want to make a living, I'm thinking that the really smart people need to stop helping billionaires and start helping themselves.
>>"ad space available -- low rates!!!"
So, all pop-sci books, no papers. Ie. nothing. That's what I figured. Your position can be trivially ignored.
9,000 times faster executing against the hardware versus the PC simulation of the same.
Vastly differing concepts. Does *ANYONE* think about this stuff before they publish utter tripe?
To the contrary. It's perfectly possible to believe in all manner of magic, a soul, a divine creator, etc. without assuming those things are necessary precursors to the existence of a mind. It could be that we create an artificial mind that lacks a soul, to whatever effect that might have. Terminators anyone?
Or from another perspective - if we accept that the brain is only the infrastructure, and that a soul must also alight upon it to give rise to a mind in the union, that does not imply that a soul could not alight upon a simulated brain to similarly give rise to a mind. I don't believe anyone has conclusively determined the exact rules governing the acquisition of a soul. Perhaps you could buy a small second-hand one from the devil and install it in your supercomputer?
--- Most topics have many sides worth arguing, allow me to take one opposite you.
They'll add so many new features nobody needed to Word it will still take as long to load a doc as it did in Window '95 and now takes in Win 8.
all discussion above ignores the inputs, where and how the nervous system is and the sensors are attached?, until there is a close approximation to types of inputs from the sensors and where they are 'attached' and how the response modifies the system, these
systems remain programs, to step outside that you need to look at how 'consciousness' developed in the first place- as part of a larger environment, we, or something like we, will eventually get there and when that happens, the issues of constructing and destructing sentient awareness will be very hard.