On-Chip Neuromorphic Hardware Will Optimize Size, Power and Speed of Computing Devicesmkw3a@virginia.edu
The Virginia NanO-computing (ViNO) group simulates and designs the hardware underlying the internet of things, from exploring the fundamental physics of emerging materials to quantum transport of electrons, to projecting overall device, circuit and system-level performance for memory, logic and sensing applications. Led by Avik Ghosh, professor of electrical and computer engineering and physics, the team develops state-of-the-art computational models and collaborates with experimentalists to understand the limits of various low-power electronic computing paradigms.
At the lowest atomistic level of modeling, the ViNO group specializes in exploring fundamental physical properties of a wide range of nanomaterials for emerging device technologies. Examples include nanomagnetic alloys that can store non-volatile data at high-bit density, 2-D materials such as graphene and topological insulators that capitalize on unconventional electron flow at very high mobility, compositionally graded thermal interface materials that minimize heat loss, and digitally grown III-V alloys and polycrystalline lead salts for high-sensitivity, single-photon detectors.
At the opposite, higher level of systems modeling, the ViNO group is looking at “on-chip” neuromorphic hardware to optimize the trade-off between a device’s size and power requirements and a software algorithm’s processing speed. The group's simulations show that a noisy low-barrier magnet may enable the design of a scalable low-power hardware unit that behaves like an analog neuron, which can be used in turn to build large-scale hardware neural networks for real-time learning and prediction.
Neural networks are computing systems that learn to perform tasks through training sequences, without being pre-programmed with task-specific rules. This resulting artificial analog neural net could potentially be attached directly on chip with an image sensor to identify and track moving objects on video in real time, and for deployment in self-driving automobiles, robots and unmanned autonomous vehicles.
A self-contained, low-power hardware neural chip might also be trained to recognize an individual’s medical signals, similar to how an electrocardiogram monitors heartbeats, or to identify atypical events and quickly classify the type of anomaly for real-time personalized medicine. Judicious on-chip processing of sensor data can greatly reduce size, weight and power of such networks, enabling the chip to operate “off line”—in the absence of a reliable wifi signal, and to protect against cyber-hacking.
The ViNO research leverages funding from NASA, the Defense Advanced Research Projects Agency, the Semiconductor Research Corporation’s Joint University Microelectronics Program, and a multi-university National Science Foundation Industry-University Cooperative Research Center on Multifunctional Integrated Systems Technology, for which Ghosh is a site-leader.