At the dawn of the computer age, sixty years ago, the ENIAC had 18,000 vacuum tubes and consumed 160,000W. In 1997, Penn students shrunk the ENIAC down to a 40mm^2 silicon chip that consumed 10W. Despite such remarkable progress, duplicating the brainís performance remains elusive. The brain uses 10W to perform 10^16 synaptic events per second (ten petaops), whereas a state-of-the-art computer uses 100W to perform 10^9 instructions per second (one gigaop). At this rate, a computer as powerful as the brain would burn 10^9W. One gigawatt!

Profligate power consumption makes it impractical to build an artificial brain or to replace damaged neural tissue with electronic prostheses.

Is silicon technology grossly inferior or are existing designs simply wasteful? Computers use transistors, while brains compute with ion channels. Electrons move through a crystal a million times faster than ions move through a liquid when driven by the same electric field. Consequently, transistors switch 25,000 times faster despite their longer channel length. But they consume 50 times more energy

than a synapse, the device neurons signal each other with using a few ion-channels. However, the transistor can match the synapses' switching energy by dropping its operating voltage to a comparable level and still be 3,600 times faster! Silicon devices are clearly superior; therefore digital computers must be prodigiously wasteful.

A computer uses a million times more stuff -- devices and associated wiring -- for each processing step, than the brain does for each synaptic event.

We are learning from the brain how to compute efficiently; our work is also yielding hypotheses about the brain's efficiency. At the device level, we exploit parallels between electrodiffusion of electrons and ions through transistor channels and ion channels. At the circuit level, we implement synaptic interaction, dendritic integration and spike generation using a few transistors. At the system level, we realize plasticity by rerouting virtual connections. We are exploring different mechanisms the brain uses to compute, including synaptic organization for vision, spike timing for audition, firing-modes for attention, and anatomical plasticity for learning.