How to replicate the squishy sophistication of the human brain in hard metal and silicon? IBM thinks it's found a way, and to prove it has built and tested two new "cognitive computing" microchips whose design is inspired by the human brain.
In the mammalian brain, neurons send chemical signals to each other across tiny gaps called synapses. A neuron's long "tail", the axon, sends the signals from its multiple terminals; the receptive parts of other neurons – the dendrites – collect them.
Each of IBM's brain-mimicking silicon chips is a few square millimetres in size and holds a grid of 256 parallel wires that represent dendrites of computational "neurons" crossed at right angles by other wires standing in for axons. The "synapses" are 45-nanometre transistors connecting the criss-crossing wires and act as the chips' memory; one chip has 262,144 of them and the other 65,536. Each electrical signal crossing a synapse consumes just 45 picajoules – a thousandth of what typical computer chips use.
Because the neurons and synapses are so close together, the pieces of hardware responsible for computation and memory are also much closer than in ordinary computer chips. Conventionally, the memory sits to the side of the processor, but in the new chips the memory – the synapses – and the processors – the neurons – are on top of each other, so they don't need to use as much energy sending electrons back and forth. That means the chips can perform parallel processing far more efficiently than conventional computers.
In preliminary tests, the chips were able to play a game of Pong, control a virtual car on a racecourse and identify an image or digit drawn on a screen. These are all tasks computers have accomplished before, but the new chips managed to complete them without needing a specialised program for each task. The chips can also "learn" how to complete each task if trained.