National lab buys brain- like IBM chips
Lawrence Livermore National Laboratory just bought a $ 1 million supercomputer that mimics the human brain to perform complex cognitive tasks while consuming less power than a lightbulb.
The national lab is the first testbed for an array of TrueNorth chips developed by IBM. TrueNorth takes a radically different approach to computing than the von Neumann architecture in use since 1946’ s ENIAC. That architecture, which still underlies current technology, relies on sequential processing and is limited by a bottleneck between processor and memory.
“The brain has a very different computational structure,” said Dharmendra Modha, IBM chief scientist for brain- inspired computing. “Our goal was to take Mother Nature’s blueprints and translate them to a computer.”
Modha, based in IBM Research’s Almaden campus in San Jose, has been working on the chip for 12 years.
The postage- stamp- size TrueNorth uses a distributed and parallel approach to process information, similar to the way brains constantly handle a tumble of sensory information. Running about 50 times faster than today’s most advanced systems, it should excel at deep
learning, a form of artificial intelligence.
In five to seven years, TrueNorth chips in mobile phones could let them recognize specific faces. Placed in smart eyeglasses, they could help blind people navigate complex environments, Modha said. The architecture could be scaled up to embed in medical imaging machines that could give immediate diagnoses. It could be used in robots or automobiles, as well as in networked servers in the cloud.
“This is the boundary of where science meets society,” he said.
Lawrence Livermore bought an array of 16 TrueNorth chips, arranged in a container the size of a laptop computer. The whole shebang runs on a mere 2.5 watts of power — a tenth of that consumed by a dim light bulb. The system packs a total of 16 million simulated neurons connected by 4 billion digital synapses.
While it may be laptop- size, it’s hardly a plug- and- play device. It will take about six months to integrate with existing computers, said Brian Van Essen, computing and informatics group head at Lawrence Livermore. IBM will participate in that process. “There’s a rich ecosystem of opensource software and tools that IBM is developing” to work with TrueNorth,” he said.
Those include a simulator, a programming language and integrated development environ- ment, and a library of algorithms, plus applications, firmware, tools for composing neural networks for deep learning, a teaching curriculum and the ability to sit in the cloud, IBM said.
Once the array is ready, the lab will put it to the test.
“We’re looking to use it on a number of problems in pattern recognition, classification and inference,” Van Essen said. It will monitor high- performance computing simulations of hydrodynamic applications — basically mixing two different fluid types. It will classify overhead-aerial images. It will “help ensure the safety, security and reliability of the nation’s nuclear deterrent system” without the need for underground testing, the lab said. It will help with additive manufacturing — 3- D printing — and with cybersecurity tasks.
“This is an important part of designing nextgeneration supercomputers,” said lab spokesman Don Johnston.