The U.S. Air Force has commissioned a “brain-inspired supercomputer” powered by IBM’s TrueNorth neuromorphic 64-chip array for potential application in aircraft and other embedded, mobile, autonomous systems with size, weight, and power constraints. With more than 64 million neurons and 16 billion synapses, the e-brain’s neurosynaptic array will perform real-time pattern recognition and sensory processing faster and more accurately than a human brain while requiring just 10 watts of power.
Jet pilots are increasingly turning to automated recognition to tell friend from foe as the speed and maneuverability of their aircraft operate at the very edge of their ability to stay conscious and alert. During dogfights, the sky fills with anti-aircraft fire, rockets, drones, flares, and decoys that can overwhelm the capabilities of human pilots to identify genuine targets. The TrueNorth e-brain array, on the other hand, can use all 64 of its chips to distinguish between the real targets and the false, and if low-power like the human brain, it can theoretically be extended beyond the 64-chip prototype to handle any number of real-time recognition tasks simultaneously.
The Air Force aims to put IBM’s e-brain theory to the test not only in aircraft but in a variety of embedded, mobile, and autonomous applications, from drones to robotics. It has not specified any limit to the number potential military applications and has implied that civilian applications will arise from its work.
“Air Force Research Lab in Rome, New York, has been the earliest adopter of the TrueNorth systems and ecosystem and is a center of excellence for neuromorphic computing through which many national and international government agencies are researching with TrueNorth,” IBM fellow Dharmendra S. Modha, chief scientist of brain-inspired computing at IBM Research-Almaden (San Jose, Calif.), told EE Times. The TrueNorth ecosystem “is now in use at 40-plus institutions and universities over five continents,” Modha said.
“IBM has increased the TrueNorth neuron count from 256 in 2011 to 64 million in 2017,” or 800 percent annually since 2011, Modha said. Over that time, he said, IBM has “demonstrated that TrueNorth is extremely efficient for mapping inference with deep networks, achieving greater than 7,000 frames/second/watt on the CIFAR-100 data set,” a Canadian Institute for Advance Research data set for object recognition. Benchmarks demonstrate 1,500-frame/second analytics using just 200 mW.
Being built at the U.S. Air Force Research Lab (AFRL), the hardware TrueNorth e-brain is intended to make the training of large, deep neural networks possible in real time, even though the 1- to 10-Hz clock speed of the prototype will enable massive chip arrays to consume less power that a single Power8 processor.
Besides target recognition, the TrueNorth e-brain’s neurosynaptic 64-chip array will deliver other artificial intelligence functions in real time, such as video analytics, speech recognition, machine-language translation, audio and text analysis, information discovery, and similar tasks destined for extension to civilian applications.
The TrueNorth e-brain’s neurosynaptic 64-chip array can accept multiple, distributed inputs and “fuse” them to perform perception (right-brain) and symbol-processing (left-brain) functions faster than conventional computers can, resulting in total situational awareness. With the ability to perform parallel data-stream processing and parallel model analytics simultaneously, the TrueNorth e-brain is intended to outperform human brains in complex combat situations where lives hang in the balance.
The entire TrueNorth e-brain neurosynaptic 64-chip array fits in a 4U (7-inch-high) rack slot, allowing eight linked systems to be installed in a standard rack. Each processor in the array uses 5.4 billion transistors, divided among 4,096 cores housing 1 million digital neurons with 256 million synapses.
Funding for TrueNorth was originally provided by the Defense Advanced Research Project Agency’s Systems of Neuromorphic Adaptive Plastic Scalable Electronics (SyNAPSE) program in collaboration with Cornell University.