Binary Synaptic Connections Based on Memory Switching in a-Si:H for Artificial Neural Networks

Abstract

AbstractNonvolatile, associative electronic memory based on neural network models promises high (−109 bits/cm2 ) information storage density since the information is stored in a matrix of only two-terminal, passive interconnections (synapses). An electronic memory switch at each interconnect would be ideal as a programmable synapse. The massive parallelism in the architecture, however, requires that the ‘ON’ state of a synaptic connection must be unusually ‘weak’ (i.e., highly resistive). For example, a binary synapse should be 106Ω in its ‘ON’ state and >109Ω in the ‘OFF’ state for a 1024 × 1024 matrix (−256 K bits of programmable read only memory, PROM). The small deliverable switching energy dictated by the resistive ‘ON’ state requirement is a new constraint for switching in thin films. Memory switching in hydrogenated amorphous silicon (a-Si:H) along with ballast (current limiting) resistors patterned from resistivity-tailored, amorphous Ge-metal alloys are investigated for a binary PROM matrix. A lμm2 area of a-Si:H could be switched from a −1010Ω (OFF state) to −105Ω (ON state) by a voltage pulse of lpsec duration, with the switching energy of −1 nanojoule, delivered through a 106Ω ballast resistor. Programmable, read-only, 1600 synapse (40x40) test arrays of uniform connection strengths (variation ±2%) and a-Si:H switching elements have been fabricated. Suitability of the memory switching in a-Si:H for high-density neural networks is discussed.