DHARMENDRA S. MODHA is probably the only microchip architect on the planet whose team includes a psychiatrist—and it's not for keeping his engineers sane. Rather his collaborators, a consortium of five universities and as many IBM labs, are working on a microchip modeled after neurons.

They call their research “cognitive computing,” and its first products, two microchips each made of 256 artificial neurons, were unveiled in August. Right now all they can do is beat visitors at Pong or navigate a simple maze. The ultimate goal, though, is ambitious: to put the neural computing power of the human brain in a small package of silicon. The program, SyNAPSE, which is funded by the U.S. Defense Advanced Research Projects Agency, is building a microprocessor with 10 billion neurons and 100 trillion synapses, roughly equivalent in scale to one hemisphere of the human brain. They expect it to be no bigger than two liters in volume and to consume as much electricity as 10 100-watt lightbulbs.

Despite appearances, Modha insists he is not trying to create a brain. Instead his team is trying to create an alternative to the architecture common to nearly every computer constructed since its invention. Ordinary chips must pass instructions and data through a single, narrow channel, which limits their top speed. In Modha's alternative, each artificial neuron will have its own channel, baking in massively parallel processing capabilities from the beginning. “What we are building is a universal substrate, a platform technology, which can serve as the basis for a wide array of applications,” Modha says.

If successful, this approach would be the culmination of 30 years of work on simulated neural networks, says Don Edwards, a neuroscientist at Georgia State University. Even IBM's competitors are impressed. “Neuromorphic processing offers the potential for solving problems that are difficult—some would say impossible—to address through conventional system designs,” says Barry Bolding, vice president of Cray, headquartered in Seattle.

Modha emphasizes that cognitive-computing architectures will not replace conventional computers but complement them, preprocessing information from the noisy real world and transforming it into symbols that conventional computers are comfortable with. For example, Modha's chips would excel at pattern recognition, like picking a face out of a crowd, then sending the person's identity to a conventional computer.

If it all sounds a little too much like the rise of the machines, perhaps it is small comfort that these chips would be bad at mathematics. “Just like a brain is inefficient to represent on today's computers, the very fast addition and subtraction that conventional computers are good at is very inefficient on a brainlike network. Neither can replace the other,” Modha says.