Scientists cram a million ‘neurons’ on postage stamp-sized computer chip
Fri 8 Aug 2014
A group of scientists has unveiled a new brain-imitating computer chip which fits over one million digital ‘neurons’ into a surface no larger than 3cm across, and uses less power than a hearing aid.
The researchers, led by IBM, described the chip as a “stamp-sized supercomputer”, with each individual neuron connecting to 256 others. Working together, these neurons are capable, for example, of identifying objects and activities in images and videos, while using nearly four times less power than a standard microprocessor.
Published in the journal of Science, the work is the culmination of a long-running project. “The cumulative total is over 200 person-years of work,” suggested Dr Dharmendra Modha, the research’s senior author.
Described by some experts as a technological breakthrough, the scientists involved hope that the cognitive chip can transform business, science, technology and society by enabling vision, audition and multi-sensory applications.
Called TrueNorth, the chip is not yet expected to be commercially useful as new software needs to be completely written from scratch to run with the chip’s “inspired” computation system.
Modha explained that TrueNorth “integrates computation, communication and memory very closely.” Instead of units of traditional binary ones and zeroes, the chip uses spikes. When an input is active enough, one of the chips’ ‘neurons’ creates a spike and sends this across to other ‘neurons.’
The framework behind the TrueNorth chip is built up from ‘neurosynaptic cores’ each containing 256 neurons – a technology first launched by IBM in 2011. The scientists then placed these core blocks into an interlinked 64-by-64 grid, generating a single 3cm chip containing more than one million neurons and 256 million ‘synapses’ or connections.
The technology is also scalable – the neurons can be connected together to provide even more computational power. Modha demonstrated this flexibility by revealing a 16-chip system with 16 million programmable neurons and four billion programmable connections.
Despite emulating the organisation of the brain – the TrueNorth chip is still a long way off matching biological neurons which pass signals to each other across as many as 1,000 trillion synaptic connections.
Demonstrating TrueNorth’s capabilities, Modha set the chip inside a video filmed from a tower at Stanford University. TrueNorth successfully analysed the moving image, identifying groups of pixels which represented pedestrians, cars, cyclists and buses.
Sophie Wilson, a Cambridge-based Royal Academy of Engineering fellow, told the BBC that it will be interesting to see how TrueNorth develops. “It’s clear that conventional scalar processing is getting very tricky for some of these tasks,” she said.
“Google Images, for example, does a marvellous job of recognising pictures of cats – but it is using large arrays of computers to do that,” she continued.
However, TrueNorth is not without rivals. A chip made by Movidius, designed especially for processing images, uses even less power than the IBM creation, also copying elements which imitate human brain functions.