About a week ago the Lawrence Livermore National Laboratory in the US bought the first chip from IBM Research which it is calling a neurosynaptic or Neuromophic computer chip meaning it is one which mimics human architectures. It is being sold as TrueNorth – and is just out. This is said to emulate human neurons and this deal involves a configuration which is the equivalent of 16 million neurons and 4 billion synapses, but it only consumes the energy equivalent of a tablet – a mere 2.5 watts of power.
This is the result of 200 collaborators working together since 2008, and IBM says that it can be installed just as easily in a smartphone as a supercomputer in the cloud, and that it plans to include its capabilities in its Watson program as soon as it can.
IBM is launching it the same way you launch any chip – with an end-to-end ecosystem complete with a new simulator, a new programming language, an integrated programming environment, new libraries and algorithms and new complete applications so that existing programmers will find it relatively easy to use.
IBM says it is working to make the ecosystem available as widely as possible to IBMers, universities, business partners, start-ups, and customers.
The architecture can solve a wide class of problems from vision, auditory, and multi-sensory fusion, and has the potential to revolutionize the computer industry by integrating brain-like capability into devices where computation is constrained by power and speed.
Systems built with True North can process high-dimensional, noisy sensory data in real time, while consuming orders of magnitude less power than conventional computer architectures. We suspect what this means is that after a bit of experimentation, in device assistants will immediately emerge using chips like TrueNorth, inside every major smartphone.
IBM listed smart phones, sensor networks, self-driving automobiles, robots, public safety, medical imaging, real-time video analysis, signal processing, olfactory detection, and digital pathology as likely uses. And in the cloud in the form of synaptic supercomputers it could be targeted at multimedia processing.
IBM says it has been working with iniLabs, which has built a retinal camera—and they have begun investigating extremely low-power end-to-end vision systems.
Effectively this is a neural network design able to infer complex cognitive tasks such as pattern recognition and integrated sensory processing, more easily than regular CPUs.
The Livermore installation will be used to help out with cyber security for US nuclear deterrents and to start with this is just a feasibility study, but it will be interesting to see where his chip leads and which of many possible segments takes the lead in building out systems using machine learning. The key to this chip is that it uses so little power. It is the result of 4 separate phases of development, each one separately funded by several $ millions, all leading to this point, since 2008.
“Neuromorphic computing opens very exciting new possibilities and is consistent with what we see as the future of the high performance computing and simulation at the heart of our national security missions,” said Jim Brase, LLNL deputy associate director for Data Science. “The potential capabilities neuromorphic computing represents and the machine intelligence that these will enable will change how we do science.”
The technology represents a fundamental departure from computer design that has been prevalent for the past 70 years and could be a powerful complement in the development of next-generation supercomputers, says IBM, and it claims it may be able to perform at exascale speeds (1 billion, billion floating point operations per second) which is 50 times faster than today’s most advanced systems.
A single TrueNorth processor consists of 5.4 billion transistors wired together to create an array of 1 million digital neurons which communicate with one another via 256 million electrical synapses. At 0.8 volts, it consumes 70 milliwatts of power running in real time and delivers 46 giga synaptic operations per second – orders of magnitude lower energy than a conventional computer running inference on the same neural network.
TrueNorth was originally developed under the auspices of Defense Advanced Research Projects Agency’s (DARPA) Systems of Neuromorphic Adaptive Plastic Scalable Electronics (SyNAPSE) program in collaboration with Cornell University.
Under terms of the contract, LLNL will receive a 16-chip TrueNorth system representing a total of 16 million neurons and 4 billion synapses. The human brain has something like 86 billion neurons for comparison.
IBM in 2011 tried first to build what it describes as “repeating canonical microcircuits” which were the equivalent of a worm-scale brain, which integrated computation and memory. Now it has shrunk the neurosynaptic core by 15-fold in area and 100-fold in power, and has tiled 4,096 cores via an on-chip network to create TrueNorth, fabricated in Samsung’s 28nm process.
Unlike the prevailing von Neumann architecture TrueNorth has a parallel, distributed, modular, scalable, fault-tolerant, flexible architecture that integrates computation, communication, and memory and has no clock.
IBM says it hasn’t exactly built a brain, but a computer which is inspired by the brain.
IBM warned, “We are not there yet. Indeed, TrueNorth is a direction and not a destination! The end goal is building intelligent business machines that enable a cognitive planet, while transforming industries.”