Jun 19, 2012
A couple of years back scientists at HP figured out how to make memristors viable. Memristors were first conceived of back in the 1970's and are components that remember (for lack of a better term) how much current passed through them for a particular interval of time. They've been compared to neurons in that the more often they fire, the more likely they are to fire in the future. On the other side of the house, scientists have been trying for decades to figure out the principles (and combination of mechanisms) by which organic brains operate. They're not binary devices, nor are they wholly analog systems, but some weird combination of things that leaves everyone scratching their heads. Some researchers decided that designing computers that mimic the functioning of brains in certain ways might figure out how organic brains work. For years they have been designing processors that, to varying extents are similar to grey matter. Alas, a few things were missing and truly neuromorphic (having similar structures to brains, roughly translated) processing hasn't come together. That is, until scientists at Intel's Circuit Research Lab announced that they'd developed a workable design for a neuromorphic chip.
Their design incorporates both memristors as well as unusual variants of transistors called spin valves (you can think of them as miniscule magnets that change their axis of rotation based upon the axes of rotation of the electrons that pass through them) in a three-dimensional latticework of microcircuitry. Judging by the rendered graphic with that article as well as reading some of their research, it seems that they worked specifically to replicate neuronal structures with their design in the hope that it would function more like a neural network than a von Neumann computer. They seem to hope to see if they can replicate certain details of organic brains' functionality with this design. Also - and this is the bit that jumped out at me - their design uses less power (and thus generates less waste heat) than conventional CPU designs deployed today. There's no word yet on the clock speed their design runs at, though I'd wager it's a lot slower than the 3.5+ GHz processor cores commonly sold today. This design is only a prototype, a proof of concept implementation, if you like so don't expect to see it for sale anytime soon.
In fact, I have my doubts that it would be useful for general purpose computation. Assuming for the moment that their neuromorphic architecture is viable for information processing, that does not immediately mean that it would be suitable for the sorts of tasks that current processor cores are applied to. By this, I mean that a neural network-like processor core may not be particularly good at arithmatic (which is all a lot of processing today is - crunching numbers (the real deal is the context that numbers are interpreted in, anyway)). They're probably not going to be good at the sort of math that is ordinarily farmed out to GPUs, either. But, like organic neural networks, I think that neuromorphic chips might be good at pattern recognition - facial recognition, log file analysis, data mining, and inferring the state of an environment from sets of many different kinds of sensors. In other words, things that are computationally expensive but possible today might one day become easier (and thus more widespread) if neuromorphic architectures pan out. Time will tell.