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."

Comments

Article and summary is misleading

[aXis100]

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.

Still a long way from brain-boxes

[Immerman]

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?

The interesting bits

[Anonymous Coward]

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.