IBM TrueNorth: 4096 core CPU

[[XC] Synthetickiller]

Saw this & figured someone should get around to posting it.

The new processor, which the team is calling TrueNorth, takes a radically different approach. Its 5.4 billion transistors include over 4,000 individual cores, each of which contain a collection of circuitry that behaves like a set of neurons. Each core has over 100,000 bits of memory, which store things like the neuron's state, the addresses of the neurons it receives signals from, and the addresses of the neurons it sends signals to. The memory also holds a value that reflects the strength of various connections, something seen in real neurons. Each core can receive input from 256 different "neurons" and can send spikes to a further 256.

The core also contains the communications hardware needed to send the spikes on to their destination. Since the whole chip is set up as a grid of neurons, addressing is as simple as providing x- and y-coordinates to get to the right core and then a neuron ID to get to the right recipient. The cores also contain random number generators to fully model the somewhat stochastic spiking activity seen in real neurons.

In total, TrueNorth has a million programmable neurons that can establish 256 million connections among themselves. The whole thing is kept moving by a clock that operates at a leisurely 1kHz. But the communications happen asynchronously, and any core that doesn't have anything to do simply sits idle. As a result, the power density of TrueNorth is 20mW per square centimeter (a typical modern processor's figure is somewhere above 50 Watts). The chip was fabricated by Samsung using a 28nm process.

Articles here:

http://www.wired.com/2014/08/ibm-unv...ocessor-cores/

http://arstechnica.com/science/2014/...icial-neurons/

Just one more step closer to building an artificial brain. At least the power consumption is great, despite the "low" clock speed of 1khz.

[PatRaceTin]

Who will buy the TrueNorth ?

[xlink]

Who will buy the TrueNorth ?

People interested in doing machine learning techniques such as artificial neural networks, perceptrons, etc.
I wonder if it would work well for other techniques such as random forests, support vector machines and logistic ensemble models.

I'm NOT an expert in this area yet. Working on it. Normally all those techniques are done in software, this sounds like it actually does it all in hardware (big speed up there) which will help in cases where data sizes get large and asymptotic scaling for any given algorithm is non-linear (think worse than n log n)

[mattkosem]

So this will be at the heart of skynet?

--Matt

[tool_462]

So this will be at the heart of skynet?

--Matt

I believe it will be the brain rather than the heart.

[mattkosem]

I believe it will be the brain rather than the heart.

Hah, indeed.

[hurleybird]

What I find fascinating is the thought of future version of these chips.

Conservatively I bet you could cram 10 million neurons on a 7nm die (linear scaling would be around 16 million, but I'm accounting for the need to let each neuron form a greater number of connections), up the frequency a few times, and stack ten or more dies on a chip.

By 2020 you could easily see a product with 100+ million neurons. Imagine what a supercomputer with a few thousand of those bad boys might do.

Skynet here we come.

[haylui]

in Hebrew www = 666 ? so this is it?