Personal computing discussed

Moderators: Flying Fox, morphine

 
Khali
Gerbil XP
Topic Author
Posts: 391
Joined: Thu Dec 20, 2012 5:31 pm

IBM unveiles "the first neurosynaptic computer" chip

Fri Aug 08, 2014 3:45 pm

Here are the highlights from the article from IBM.

5.4 billion transistors.
one of the largest CMOS chips ever built.
only consumes 70mW.
one million programmable neurons, 256 million programmable synapses and 46 billion synaptic operations per second per watt.

http://www-03.ibm.com/press/us/en/pressrelease/44529.wss
 
chuckula
Gold subscriber
Graphmaster Gerbil
Posts: 1329
Joined: Wed Jan 23, 2008 9:18 pm
Location: Probably where I don't belong.

Re: IBM unveiles "the first neurosynaptic computer" chip

Fri Aug 08, 2014 3:56 pm

Whoa whoa whoa... let's step back some of IBM's marketing BS just a notch there: If you read what IBM actually wrote, they claim that their magical chip consumes 70 milliwatts running at what they refer to as "biological realtime". Now... WTF does "biological realtime" mean? Well, using a relatively sane definition of the firing rates of neurons in the human brain, the neurons fire at anything between ~ 1 Hz up to 200 Hz depending on the neuron type and level of brain activity (see: http://neuroblog.stanford.edu/?p=4541 ).

So... 70 Milliwatts at maybe (optimistically) 200 Hz?? Not too shabby, but nothing miraculous either especially if IBM is being smart enough to use power gating so those "5.4 Billion" transistors look more like the few hundred million transistors in your smartphone running at any one moment in time... oh and those transistors are running at an intentionally crippled clock rate.
4770K @ 4.7 GHz; 32GB DDR3-2133; GTX-1080; 512GB 840 Pro (2x); Fractal Define XL-R2; NZXT Kraken-X60
--Many thanks to the TR Forum for advice in getting it built.
 
Kougar
Silver subscriber
Gerbil Jedi
Posts: 1704
Joined: Tue Dec 02, 2008 2:12 am
Location: Texas

Re: IBM unveiles "the first neurosynaptic computer" chip

Sun Aug 10, 2014 4:20 am

It's cool, but yeah it's nowhere close to a brain. The human variety has 100 billion neurons and 100 trillion synapses (versus their 256 million) so they have quite a ways to go... :P
 
Igor_Kavinski
Gerbil XP
Posts: 399
Joined: Fri Dec 22, 2006 2:34 am

Re: IBM unveiles "the first neurosynaptic computer" chip

Sun Aug 10, 2014 6:50 am

Just a few days now before Terminator arrives and destroys IBM...
 
vargis14
Gerbil Jedi
Posts: 1899
Joined: Fri Aug 20, 2010 6:03 pm
Location: philly suburbs

Re: IBM unveiles "the first neurosynaptic computer" chip

Sun Aug 10, 2014 7:51 am

Wonder what they will name the finished project? Lucy like the movie that is out now? Quote"I can see and feel everything"
2600k@4848mhz @1.4v CM Nepton40XL 16gb Ram 2x EVGA GTX770 4gb Classified cards in SLI@1280mhz Stock boost on a GAP67-UD4-B3, SBlaster Z powered by TX-850 PSU pushing a 34" LG 21/9 3440-1440 IPS panel. Pieced together 2.1 sound system
 
the
Gold subscriber
Gerbil Elite
Posts: 767
Joined: Tue Jun 29, 2010 2:26 am

Re: IBM unveiles "the first neurosynaptic computer" chip

Fri Aug 15, 2014 12:58 am

Kougar wrote:
It's cool, but yeah it's nowhere close to a brain. The human variety has 100 billion neurons and 100 trillion synapses (versus their 256 million) so they have quite a ways to go... :P


I wonder how well this would scale across multiple sockets. The low base clock speed could be tunneled over vastly higher speed serial links. To get to the human level that'd require roughly a quarter million chips to be tethered together. That sounds like a lot (and well it still is a lot) but IBM has managed to get about half of that clustered together in their big BlueGene/Q super computers. If they could get the interconnect to work, and that's a big if, they'd be able to approach the human level of synapses in a few years.
Dual Opteron 6376, 96 GB DDR3, Asus KGPE-D16, GTX 970
Mac Pro Dual Xeon E5645, 48 GB DDR3, GTX 770
Core i7 3930K@4.2 Ghz, 32 GB DDR3, GA-X79-UP5-Wifi
Core i7 2600K@4.4 Ghz, 16 GB DDR3, GTX 970, GA-X68XP-UD4
 
just brew it!
Gold subscriber
Administrator
Posts: 47576
Joined: Tue Aug 20, 2002 10:51 pm
Location: Somewhere, having a beer

Re: IBM unveiles "the first neurosynaptic computer" chip

Fri Aug 15, 2014 1:14 am

the wrote:
I wonder how well this would scale across multiple sockets.

Probably not well. The power of neural nets derives in part from their ability to form lots of arbitrary interconnections which can all be active at once. Even with multiple high-speed links I think you'd quickly saturate the communications fabric.
If the world isn't making sense to you, you're either drinking too much or not drinking enough.
 
jihadjoe
Gerbil Elite
Posts: 571
Joined: Mon Dec 06, 2010 11:34 am

Re: IBM unveiles "the first neurosynaptic computer" chip

Fri Aug 15, 2014 5:53 am

What I read:

46 giga-synaptic ops per sec (GSOPS i guess?) / W
* .07W
= 3.22 GSOPS

Sounds like a lazy bunch of neurons. Maybe they should use caffeine for thermal paste.
 
the
Gold subscriber
Gerbil Elite
Posts: 767
Joined: Tue Jun 29, 2010 2:26 am

Re: IBM unveiles "the first neurosynaptic computer" chip

Fri Aug 15, 2014 2:09 pm

just brew it! wrote:
the wrote:
I wonder how well this would scale across multiple sockets.

Probably not well. The power of neural nets derives in part from their ability to form lots of arbitrary interconnections which can all be active at once. Even with multiple high-speed links I think you'd quickly saturate the communications fabric.


I think on smaller number of dies (say <16) scaling shouldn't be that much of an issue if the low clock frequencies are correct. IBM is known for producing high clock inter-processor links so multiplexing several slower connections into one high speed bus shouldn't be a problem.

That still leaves much scaling to get to the human level. Still, with newer chips and better interconnects (silicon photonics?), it is conceptually feasible. Going for it would be dependent on how many 'if's' I mentioned get explored with IBM's current chip.
Dual Opteron 6376, 96 GB DDR3, Asus KGPE-D16, GTX 970
Mac Pro Dual Xeon E5645, 48 GB DDR3, GTX 770
Core i7 3930K@4.2 Ghz, 32 GB DDR3, GA-X79-UP5-Wifi
Core i7 2600K@4.4 Ghz, 16 GB DDR3, GTX 970, GA-X68XP-UD4
 
just brew it!
Gold subscriber
Administrator
Posts: 47576
Joined: Tue Aug 20, 2002 10:51 pm
Location: Somewhere, having a beer

Re: IBM unveiles "the first neurosynaptic computer" chip

Fri Aug 15, 2014 2:50 pm

I still think you're going to get into trouble. Within each chip it is a many-to-many topology between neurons. Maintaining a comparable level of connectivity to neurons on other chips is going to require a huge amount of communications bandwidth (in spite of the slowness of individual neurons), and a lot of additional logic and on-chip interconnects to route those links.

Maybe you could do something with chip stacking... still think it is gonna be tough.
If the world isn't making sense to you, you're either drinking too much or not drinking enough.
 
the
Gold subscriber
Gerbil Elite
Posts: 767
Joined: Tue Jun 29, 2010 2:26 am

Re: IBM unveiles "the first neurosynaptic computer" chip

Fri Aug 15, 2014 4:19 pm

The one thing in our favor is that not all the nodes are communicating with each other simultaneously. The dreaded 0.5*n*(n-1) mesh scaling can be avoided (actually does this even merit a true mesh to mimic a human brain?). As such, this effectively becomes a networking problem and I classify that as a 'solved problem'. Bandwidth isn't free but it can be managed efficiently.

Chip stacking is an interesting idea, especially if TSV is used to enable vertical movement in the middle of a die. The catch would be keeping the middle of the stack cool but that's no different than any other massively stacked device. IBM's claims of ultra low power may enable this.

Another thought occurred to me: SMT. Essentially double the clock rate and evenly divide up the resources. Bandwidth requirements scale up to go between sockets but then you won't necessarily have to move off-die as often. It'd require a massive change to each 'core' though and it is already a massive die. A small change like SMT on a per core basis would be massively amplified by the sheer number on the silicon. This is something that'd require a die shrink or two I'd fathom.
Dual Opteron 6376, 96 GB DDR3, Asus KGPE-D16, GTX 970
Mac Pro Dual Xeon E5645, 48 GB DDR3, GTX 770
Core i7 3930K@4.2 Ghz, 32 GB DDR3, GA-X79-UP5-Wifi
Core i7 2600K@4.4 Ghz, 16 GB DDR3, GTX 970, GA-X68XP-UD4

Who is online

Users browsing this forum: No registered users and 1 guest