“This will be the first time a computing architecture is built entirely with biomimetic memristors, using no transistors or other common types of active devices,” said HRL Principal Investigator (PI) Wei Yi.
“In this program we aim to develop efficient neuromorphic circuit architectures based on memristors instead of complementary metal-oxide semiconductor (CMOS) transistors,” said co-PI Jose Cruz-Albrecht. “Neuromorphic processors that use just CMOS devices have difficulty efficiently merging memory and computation with very compact circuitry. Most require several transistors to implement just the memory portion of a single synapse, additional transistors for the most basic synaptic computation, and still more transistors for complex synaptic computations that include adaptation mechanisms. In this program we plan to develop all-memristor circuit architectures that efficiently merge memory and computation. An electronic synapse that includes memory, computation, and adaptation could be designed with as little as one memristor, which could enable more compact and energy-efficient neuromorphic processors.”
HRL will utilize its extensive integrated circuit (IC) experience to build the first-ever all-memristor processor. A crucial step will be to realize integrated device structures, IC layout, and fabrication processes of integrated active memristor electronic neurons. The goal is a transistor-free, spiking neural network processor, which operates like part of a brain, with the main computations done by passive and active memristors emulating synapses and neurons, respectively.
“Traditional von Neumann computer architecture scaling is slowing down, based on the physical limits of how many transistors will fit on a given size of silicon chip. Thus, the accepted industry wisdom is that we are finally approaching the end of what we commonly refer to as Moore’s Law,” said HRL researcher Dana Wheeler. “We’ve known for some time that if you keep cramming components onto a chip and making it faster, eventually it will get hot enough to melt the circuit. With this program, DARPA is interested in research efforts that are exploring architectures beyond Moore’s Law – forging a path suggested by Moore himself.”
“Our approach is completely novel in the sense that memristors can perform very complex computational tasks that von Neumann computers are incapable of without enormous size and huge power usage,” said Yi. “The rationale is that if we want to build an energy-saving intelligent machine, we need to learn from nature by analyzing ourselves. Human brains consume only about 20 watts of power, but can perform some tasks even a supercomputer cannot. In the brain, computations are done by neurons, which handle signal generation and processing, and synapses, which are plastic connections between neurons that remember data and adapt new information to form learning. These connections also constantly evolve and some may disappear as a result. An adaptive neural network that can mimic evolution of neuron connectivity does not yet exist, but active and passive memristors can emulate neuron and synapse activities, respectively, and thus form self-sufficient building blocks that may eventually be augmented and scaled up to build an entire electronic brain of the same size, energy efficiency, and speed of a biological brain.”
HRL Laboratories, LLC, Malibu, California (hrl.com) is a corporate research-and-development laboratory owned by The Boeing Company and General Motors specializing in research into sensors and materials, information and systems sciences, applied electromagnetics, and microelectronics. HRL provides custom research and development and performs additional R&D contract services for its LLC member companies, the U.S. government, and other commercial companies.
Media Inquiries: Michele Durant, (310) 317-5321