Skip to main content
eScholarship
Open Access Publications from the University of California

Spike Neuromorphic Carbon Nanotube Circuits

  • Author(s): Kim, Kyunghyun
  • Advisor(s): Chen, Yong
  • et al.
Abstract

The brain has a superior computation performance in comparison with supercomputers in many aspects while the brain consumes much less power (~ 20 W) than supercomputers (~ 106 W). This is mainly owing to synapses, which are the fundamental elements in the brain, and their properties such as spatiotemporal signal processing, memory, and learning. Therefore, it is crucial to implement an electronic device which can emulate the elementary functions of biological synapses in order to mimic the functions of the brain. In this research, two types of synaptic transistors (synapstors) using carbon nanotubes (CNTs) are demonstrated. First, a CNT network transistor with a poly(ethylene glycol) monomethyl ether (PEG) layer in the gate is presented. It has successfully emulated the elementary functions of biological synapses with low power consumption (250 pW/synapstor). Second, a CNT network transistor with C60 molecules is shown to mimic the essential functions of biological synapses with low power consumption (2.6 nW/synapstor). A spike neuromorphic circuit (SNC) was developed by integrating CNT synapstors. The SNC has a capability of parallel signal processing, spatiotemporal correlation, learning with low power consumption. It has both excitatory and inhibitory synapses and can generate output spikes from the accumulated post-synaptic currents. The large-scale SNC with 16,384 synapstors and 16 neurons has been designed and fabricated. The power consumption of a large-scale SNC is ~ 1.8 mW. The functions of a SNC were demonstrated. The toy drone was used as a platform to interact with the SNC. The SNC dynamically processed the sensing signals from the drone and triggered actuation of the drone in real-time. The performance of the drone was improved via the learning in SNC. The SNC has a potential to have higher signal processing speed and be more efficient in power consumption than a supercomputer when the dimension of parallel signal processing exceeds ~109.

Main Content
Current View