Gallery

Whatever happened to Steve Furber?

Posted on 16 Jun 2010 at 18:19

Steve Furber plans to model the brain with a million of the ARM processors that first appeared in Acorn computers

Acorn may have ended in honourable failure, but it’s a little-known fact that one of the company’s innovations has been spun off into one of the biggest British tech success stories of the past decade: processor firm ARM.

The company’s low-powered processors are today found in everything from the iPhone to car alarm systems, but in the early 1980s it was Acorn’s Steve Furber and Sophie Wilson who first developed the ARM architecture for use inside the Acorn Archimedes.

Nearly 30 years later, and Furber is still working with ARM processors, although on a much grander scale than the humble Archimedes. His latest project, Spinnaker, uses no fewer than a million ARM processors.

If you take the world’s biggest supercomputer at the moment, which is about one and a half petaflops, that probably is still only about 10% of what you need to do a whole brain

“Spinnaker is a machine whose architecture is inspired by the very high levels of connectivity we see in the brain, and we’re trying to reproduce that in a computer,” Furber told us. “A major application for the machine will be modelling parts of the brain and trying to understand the principles at work inside the brain.”

Furber, who is now a professor of computer engineering at the University of Manchester, hopes Spinnaker will eventually allow neuro-scientists and psychologists to test theories on how the brain operates.

Yet even with a million processors, it doesn’t even come close to matching the old grey matter. “With a million ARM processors we can get to about 1% of the full human brain,” Furber admits.

“The brain has 10¹¹ neurons running in parallel with 10¹⁵ connections where most of the action is. Estimates as to what level of computer power you’d need to model the brain vary, but it’s certainly well above the petaflop [10¹⁵ flops].

If you take the world’s biggest supercomputer at the moment, which is about one and a half petaflops, that probably is still only about 10% of what you need to do a whole brain. We need to get to exaflop [10¹⁸ flops] machines before we can comfortably claim to have the computer power to model a complete brain [in real-time].”

Furber’s intricate understanding of the ARM architecture made it a natural choice to power Spinnaker, but he insists there are more practical reasons for choosing ARM, too. “With a million ARMs it’s looking like a 50kW machine,” he said. “The electricity bill is non-trivial. If we did it with high-end processors, the electricity bill would go up by an order of magnitude.”

Does Furber, who was designing processors with a power consumption of less than a watt in the Acorn era, despair at the power demands of today’s high-end processors? “It’s a consequence of the fact that we still don’t have a good grip on building parallel software, except in some special and important cases,” he insisted.

“All supercomputers are massively parallel now, and they run applications where somebody spent a lot of time and effort to make sure they run well. For general desktop computing, we haven’t cracked that problem. Therefore, it’s economical to build a processor that works very hard and somewhat inefficiently to make a single thread go fast. The performance they get on a single thread justifies the rather silly power figures.”

Author: Barry Collins