SpiNNaker set to receive new 18-core SoC to help reverse engineer the human brain.
A British research project working towards the Grand Challenge of reverse engineering the human brain is about to move into its second generation of silicon. The work will, for now, allow biologists and psychologists to work on modelling the brains of some lower-order animals.
The SpiNNaker (Spiking Neural Network Architecture) project, led out of the University of Manchester, is set to receive its first 18-core, ARM processor-based system-on-chip from foundry UMC in the next fortnight.
Professor Steve Furber, one of the pioneering ARM team and ICL Professor of Computer Engineering at Manchester, told the Design Automation and Test in Europe (DATE) conference in Grenoble, France that the 100 million-gate device would allow for the more detailed modelling of spiking neurons in real-time.
The new SpiNNaker chip replaces a two-core prototype that the team has been using for a little under two years. The chips are combined on boards in multiple configurations – the prototype uses four.
Other partners on the project include the Universities of Southampton, Sheffield and Cambridge, and the Engineering and Physical Sciences Research Council as well as companies ARM, Silistix and Thales.
SpiNNaker has a mesh architecture where neurons are modelled to each individual processor. The main innovation is the interconnect fabric. Each chip has a router with a look-up table (LUT) that dynamically decides which processor or processors to direct a spike to.
To full mimic the human brain would require LUTs with millions of entries, but the team has optimised the table for the current version so that it requires only 1,000 to emulate realistic biological networks.
There is still a long way to go, Furber said. Having the computing power to model systems is one thing. “But we don’t know what to model,” he continued.
The longer-term goal is to build machines with 1.6 million ARM cores, but this is still some way from full simulation. By way of comparison, Furber said that 1 million ARM cores would still represent only one per cent of the human brain in terms of modeling capability. Human modelling is about a decade away.
“This has been a very large project by academic standards and I’m hopeful that it will produce interesting results very soon,” Furber said.