Massively Parallel Integrated CMOS-RRAM Arrays with Embedded Online Learning for AI on the Edge

Abstract

This project aims at transformative advances in the efficacy and efficiency of neuromorphic hardware with embedded online learning for autonomous AI on the edge. Building on foundational advances in previous research by the team on large-scale spike-based neuromorphic systems with persistent online deep learning as part of the Office of Naval Research Science of Artificial Intelligence Program, this project will deliver their practical realization at scale in custom-integrated microchips with real-time learning functionality embedded within compute-in-memory arrays. Record levels in high-density, low-energy operation of massively parallel neuromorphic hardware developed in this research will be attained by custom fabrication and experimental characterization of vertically integrated complementary metal-oxide semiconductor (CMOS) neuron arrays and emerging technology bulk resistive random-access memory (RRAM)crossbar synapse arrays. The custom high-density hybridized integration ofRRAM synapse arrays at very fine pitch (down to 200nm) on top of CMOS wafers in advanced (130nm to 180nm) CMOS technology will provide the first practical realization, at record high integration density (up to 1 billion weights on-chip) and efficiency (up to 1,000 TOPS/W, tera-operations per second per Watt), of a complete, self-contained, auto-adaptive neuromorphic compute-in-memory system for high-performance AI on the edge.

Document Details

Document Type
DoD Grant Award
Publication Date
Mar 08, 2024
Source ID
N000142412127

Entities

People

  • Gert Cauwenberghs

Organizations

  • Office of Naval Research
  • United States Navy
  • University of California, San Diego

Tags

Readers

  • Agent-Based Social Robotics and Mobile-Assisted Learning in Virtual Environments.
  • Integrated Circuit Design and Technology.

Technology Areas

  • AI & ML
  • AI & ML - DoD AI Strategy
  • AI & ML - Neural Networks
  • Microelectronics