Low-Power Chips to Model a Billion Neurons

The Brain Maker

Steve Furber
Photo: University of Manchester

In the late 1970s, IEEE Fellow Steve Furber was a mild-mannered research fellow in aerodynamics at the University of Cambridge when he was snatched up by fledgling start‑up Acorn Computers. There, in just a few years, he codesigned Britain’s popular BBC Microcomputer System and the 32-bit ARM microprocessor, which debuted in 1985 and now forms the core logic of nearly every single mobile phone on the planet.

And he did all that before he realized what he really wanted to do with his life.

Furber left Acorn in 1990 for a professorship at the University of Manchester. He watched, delightedly, as ARM rose to dominate the low-power processor market. But he wasn’t satisfied. “Processors were getting so much more efficient and powerful, but they still struggled to do things that 2-year-olds find very straightforward,” he recalls.

For example, human brains excel at pattern recognition. We can recognize a face in shadow or in direct sun, in profile or from the front, and in blurred or cutoff photos. But in computer circuitry, matching an input pattern with a stored one is notoriously brittle. Unless there is an exact match, the memory will come up empty. In the late 1990s, Furber set out to see if he could do better. “But every time I tried to resolve it in circuitry, I ended up designing a match detector that looked like a neuron,” Furber says. So he decided to devote himself full-time to brain modeling.

He used as his starting point the low-power ARM processor. By designing a novel 18-core chip and communication strategies to connect tens of thousands of those chips, he and his colleagues came up with a new kind of digital computer, called SpiNNaker, for Spiking Neural Network Architecture. As Furber writes in this issue, the massively parallel machine they are now building combines some of the best features of analog chips and digital supercomputers.

SpiNNaker’s modest goal is to model neurons flexibly and in more or less real time. But Furber believes that models like SpiNNaker will one day help shed light on the big, profound issues: the origins and nature of self-awareness and consciousness. “I don’t buy the argument that a brain can never understand itself,” he says.