Brains! —

IBM’s brain-inspired chip finds a home at Livermore National Lab

A cluster of True North processors will mimic 16 million neurons.

A cluster of True North chips.
A cluster of True North chips.

A lot of the science we cover at Ars focuses on technology development, where the risks involved mean that the early success represented by a publication doesn't typically lead to an actual product. So it's nice to be able to report on an exception to this rule. IBM's experimental neural processor, True North, has found a home at Lawrence Livermore National Lab.

True North is a radically energy-efficient design with a circuitry designed to mimic the structure of the neural connections within an animal's brain. Each chip is a big cluster of small cores that can potentially communicate with any other core on the chip. Each of these cores has its own memory and communication hardware; the memory holds information on the other cores it communicates with and how strong those connections are. The communications then take the form of a series of "spikes," bursts of activity that carry information based on their frequency and strength.

The radically different design allows the chip to get work done despite a ludicrously low clock rate: just one kiloHertz. The trade-off is that it can only host neural network software—it was designed to be compatible with any networks developed for the popular Compass neural network software package. And compared to running Compass on a traditional processor, True North used 176,000-fold less energy.

IBM also designed it to scale, as each "neuron" on the chip had the capacity to communicate with a neuron on an entirely separate chip. "We have begun building neurosynaptic super-computers," its developers wrote at the time, "by tiling multiple TrueNorth chips, creating systems with hundreds of thousands of cores, hundreds of millions of neurons, and hundreds of billion of synapses."

And that's precisely what Lawrence Livermore is getting. The hardware is a cluster of 16 True North chips, able to deploy up to 16 million silicon neurons and establish 4 billion synapses. (For context, the human brain has 86 billion neurons, so we've still got a bit of a computational lead at the moment.) When running flat out, the entire cluster will consume a grand total of 2.5 watts.

At the moment, Lawrence Livermore plans on testing whether neural networks are useful for the sorts of problems that the lab typically needs to solve. The deal, worth $1 million for IBM, includes all the support and development tools needed to implement neural networks on the cluster. But it will undoubtedly pay for itself in lower electrical bills if researchers can move any of their problems off Lawrence Livermore's more traditional supercomputers.

Channel Ars Technica