IBM, US Air Force Are Building a Neuromorphic Supercomputer

IBM and the US Air Force have announced that they’re teaming up to build a unique supercomputer based on IBM’s TrueNorth neuromorphic architecture. The new supercomputer will consist of 64 million neurons and 16 billion synapses, while using just 10W of wall power.

It’s common, when discussing CPUs, to compare them with the human brain. Superficially, the two seem similar — brains, like CPUs, receive inputs, perform calculations based on those inputs, and then return a result. But while brains and conventional CPUs may seem similar at an extremely high level, that similarity disintegrates as soon as you start examining either system in any detail.

Transistors are binary (they’re either on or off), and they can only change the behavior of other transistors that they’re directly connected to. Neurons, in contrast, have both an analog and a binary aspect. The dendrites — the receiving arms of a neuron — have analog function in that they give a little electrical ripple called a “graded potential” whenever they get a ping from an upstream neuron. If they send enough graded potentials to the cell body, the latter then sends a binary off/on pulse train down the axon. The axon of a nerve sort of “speaks” in binary, even though it has to accommodate both binary and analog input.

Scientists have spent decades creating software models that more closely resemble the way brains process information. But there’s an enormous efficiency gap when attempting to simulate something as different as a brain on modern silicon. While modern CPUs may be millions of times faster at certain calculations than any human, the human brain’s power efficiency is orders of magnitude better than the most efficient conventional CPU we can build.

IBM’s TrueNorth project is an attempt to build a neuromorphic, or brain-like CPU directly in hardware. The goal is to design superior neural nets and create artificial intelligence in power envelopes that could conceivably operate outside of data centers or fixed installations.

IBM is claiming that the TrueNorth Neurosynaptic System (that’s the official moniker) can convert and process data from multiple sources in parallel, while simultaneously pairing with more conventional processors to analyze the data.

The US Air Force Research Laboratory (AFRL) was the earliest adopter of TrueNorth for converting data into decisions, said Daniel S. Goddard, director, information directorate, U.S. Air Force Research Lab. “The new neurosynaptic system will be used to enable new computing capabilities important to AFRL’s mission to explore, prototype, and demonstrate high-impact, game-changing technologies that enable the Air Force and the nation to maintain its superior technical advantage,” he said in a statement.

The new system will fit in a 4U standard server rack with 512 million neurons in total per rack. IBM claims this represents an 800-percent annual increase over the last six years, as the first systems contained just 256 neurons.



ExtremeTech