As a part of the U.S. Department of Energy’s Advanced Scientific Computing Research program, Intel today inked a three-year agreement with Sandia National Laboratories to explore the value of neuromorphic computing for scaled-up AI problems. Sandia will kick off its work using the 50-million-neuron Loihi-based system recently delivered to its facility in Albuquerque, New Mexico. As the collaboration progresses, Intel says the labs will receive systems built on the company’s next-generation neuromorphic architecture.
Along with Intel, researchers at IBM, HP, MIT, Purdue, and Stanford hope to leverage neuromorphic computing — circuits that mimic the nervous system’s biology — to develop supercomputers 1,000 times more powerful than any today. Chips like Loihi excel at constraint satisfaction problems, which require evaluating a large number of potential solutions to identify the one or few that satisfy specific constraints. They’ve also been shown to rapidly identify the shortest paths in graphs and perform approximate image searches, as well as mathematically optimizing specific objectives over time in real-world optimization problems.
Intel’s 14-nanometer Loihi chip contains over 2 billion transistors, 130,000 artificial neurons, and 130 million synapses. Uniquely, the chip features a programmable microcode engine for on-die training of asynchronous spiking neural networks (SNNs), or AI models that incorporate time into their operating model such that the components of the model don’t process input data simultaneously. Loihi processes information up to 1,000 times faster and 10,000 more efficiently than traditional processors, and it can solve certain types of optimization problems with gains in speed and energy efficiency greater than three orders of magnitude, according to Intel. Moreover, Loihi maintains real-time performance results and uses only 30% more power when scaled up 50 times, whereas traditional hardware uses 500% more power to do the same.
Intel and Sandia hope to apply neuromorphic computing to workloads in scientific