An artist's impression of a tiled network of TrueNorth chips

Super-efficient brain mimicking chip unveiled by IBM

A super-efficient chip that mimics the way the human brain operates could revolutionise sensory data processing and provide a path to exascale computing.

IBM’s TrueNorth chip is roughly the size of a postage stamp and features 5.4 billion transistors arranged into 4,096 neurosynaptic cores that allow it to run at biological real-time – approximately the speed at which the human brain processes information – on just 70mW of power.

The chip was inspired by insights from neuroscience and has been modelled on the human cerebral cortex by creating in silico equivalents of neurons and the synapses that connect them, as detailed in an paper written in collaboration with Cornell Tech in the USA that has made the cover of tomorrow’s edition of journal Science.

The company hopes the DARPA-funded project will enable computers to emulate the brain’s capacity to process vast amounts of sensory information with very little power while taking up minimal space.

As well as overcoming energy consumption challenges limiting current supercomputer technology, the chip could open the door to complex data processing at the point of collection for applications as varied as visual, audio and even olfactory sensing.

“Imagine bringing low-power processing and putting it in a camera. We can move computation to the data,” said IBM’s chief scientist for brain-inspired computing Dharmendra Modha. “The sensor becomes the computer.”

The TrueNorth chips follows on from an initial single core hardware prototype in 2011 and a dedicated software ecosystem with a new programming language and chip simulator released in 2013 to enable developers to create applications optimised for the kind of parallel computing the device employs.

IBM has abandoned the prevailing von Neumann architecture used almost universally since 1946, which separates memory and computation, as it says its sequential processing model is incompatible with the parallel and event-driven computation of the human brain and also leads to high power consumption.

Each core integrates memory on-chip, along with computation and communication, allowing the creation of a distributed architecture that supports super-fast parallel processing as well as high fault tolerance, as multiple cores reduce the impact of individual components failing.

The one million programmable ‘spiking’ neurons and 256 million programmable synapses that make up the cores have been created using CMOS technology via Samsung’s 28nm process. The neurons mimic their biological counterparts by firing short sharp impulses before becoming inactive again, allowing the chip to be event-driven rather than constantly clocking, reducing power consumption.

The design also allows multiple TrueNorth chips to seamlessly connect to adjacent chips, which the firm hopes could pave the way for the development of future neurosynaptic supercomputers.

“We are not trying to replace the von Neuman architecture. We think of it as a left brain machine for fast, sequential number crunching, while our new architecture is a right brain machine for things like sensory processing and pattern recognition,” said Modha. “We envision hybrid computers that bring both together.”

Leading neuroscientist Dr James Olds, director of George Mason University's Krasnow Institute, is familiar with the project and says the major innovation in the TrueNorth chip has been the drastic increase in both space and power efficiency.

“Really the most impressive thing the human brain does, is to do what it does on 20 watts of electricity,” he said. “If they can get a fraction of that efficiency deployed in high performance computing it’s going to be worth a lot of money to IBM.”

In contrast to projects such as the Human Brain Project, which has come under fire from elements of the neuroscience community for attempting to create an overly ambitious in silico replica of the human brain, Olds says IBM has taken the pragmatic approach of reverse engineering those features of human cognition they see as applicable to real-world problems.

“What they’ve done is taken what they see as the fundamentally most important aspects of the biological brain architecture such as action potential spikes (in neurons), or the massively parallel construction, because they are betting on the fact that in the end that will be the only way to deliver exascale computing for a reasonable amount of money,” he said.

The ability to carry out brain-like functions using very little power could also have major implications for autonomous systems such as self-driving cars, space robotics and deep-sea drones. To test TrueNorth’s applicability to real world problems the researchers challenged it to an object detection and classification task using video of people, bicyclists, cars, trucks, and buses, with the chip achieving state-of-the-art performance.

“A key use will be image comprehension, which is something the human brain does that is really challenging for digital computers,” said OIds.

“The more you can deploy in silico solutions which mimic the human brain’s ability for image comprehension, the more powerful those machines can be in environments where they have to be utterly autonomous like the bottom of the ocean or on Mars.”

Recent articles

Info Message

Our sites use cookies to support some functionality, and to collect anonymous user data.

Learn more about IET cookies and how to control them