Copyright © 2020 Albuquerque Journal
Sandia National Laboratories and Intel Corp. are teaming up to build next-generation supercomputers that could function like human brains, instantaneously resolving complex problems with a fraction of the energy needed by today’s state-of-the-art machines.
The two entities announced a collaborative, three-year research agreement Friday that will allow Sandia scientists to study and further develop “neuromorphic computing,” an emerging science that mimics the way living brains function to process information. Intel has developed new chips and servers that enable that new computing method, and it’s now providing Sandia with advanced neuromorphic servers for laboratory use.
Intel’s newly-built neuromorphic computing chips, called “Loihi,” can process information up to 1,000 times faster and 10,000 times more efficiently in terms of energy use than traditional processors, according to the company. Through further research and development, Intel believes neuromorphic processors could provide the artificial intelligence needed to power emerging technologies in everything from advanced robotics and self-driving cars to new sensors that can sniff out dangerous chemicals or identify potential terrorist threats when sifting through massive mounds of information.
Under the new agreement, Sandia will apply Intel’s technology to complex scientific and engineering challenges, testing and evaluating the advantages of neuromorphic computing compared with today’s supercomputers. And, working together, the two entities will develop new algorithms and circuitry needed to potentially scale up neuromorphic processors to possibly 10 times the computational capacity of Intel’s current systems, said Craig Vineyard, a computer scientist and principal Sandia technical staff member who is leading the project.
“We’ll look at how neuromorphic computing can benefit and advance the science of computing,” Vineyard told the Journal. “We use supercomputers to analyze physics and chemical codes, or how molecules react or adjust to forces. We’ll explore how neuromorphic processors can be used for that type of scientific computing, and how, at large-scale, they can aid in complex analytics.”
The research could lead to new neuromorphic tools, algorithms and systems, said Mike Davies, director of Intel’s Neuromorphic Computing Lab.
“As high demand and evolving workloads become increasingly important for our national security, Intel’s collaboration with Sandia will provide the tools to successively scale neuromorphic computing solutions to an unprecedented level,” Davies said in a statement. “Sandia’s initial work will lay the foundation for the later phase of our collaboration, which will include prototyping the software, algorithms and architecture in support of next-generation, large-scale neuromorphic research systems.”
In essence, the new computing method simulates the action of brain neurons, which instantaneously react to environmental stimulation, communicating with one another to identify and analyze things. Neuromorphic computers use artificial neurons to act in the same way, selectively triggering sets of neurons for analysis and problem-solving based on particular stimuli.
In contrast, today’s computers are based on a linear computational process driven by graphic processing units, or chips, that are programmed to do specific tasks. For complex functions or calculations, those chips work in combination with each other, like individual building blocks that all come together to supply the system’s full computational ability.
In neuromorphic computing, only those artificial neurons needed for a particular task are stimulated, greatly speeding analysis and reducing the amount of energy needed for processing, Vineyard said.
“Traditional supercomputers use immense amounts of power, making them very expensive to operate,” Vineyard said. “They require liquid cooling and dedicated power lines that supply megawatts of electricity, because the entire system is turned on for every computation. Neural-driven computers don’t do that, so they use much less energy.”
In addition, unlike in traditional processors, memory and computing elements in neural-based systems are intertwined rather than separate, minimizing the distance data must travel, which reduces energy consumption and speeds computations, according to Intel. And, the neuromorphic method allows processors to learn and memorize things with much less data to grow smarter at a much faster pace.
Intel’s latest neuromorphic system encompasses 100 million artificial neurons, roughly equivalent to the brain of a small mammal. The company made that system available over the cloud early this year for members of the Intel Neuromorphic Research Community — which includes dozens of universities, research institutions and private companies — to experiment with it.
Intel has now provided Sandia with a system encompassing 50 million artificial neurons. It’s the first in a series of test beds that Intel will provide over the next three years, potentially scaling up to a system with 1 billion neurons or more, Vineyard said.
“Even at 50 million neurons, this system is already one of the five largest neuromorphic computing processors in the world,” Vineyard said. “Intel is a world-class chip manufacturer that’s building us world-class, large-scale systems. We’re excited about it.”