Neuromorphic Computing Using Ferroelectric Based Devices


In IBM Research Zurich, we develop devices for neuromorphic systems. In artificial neural networks accelerators, the “synaptic weights” are encoded in the conductance of programmable resistors. The online training of artificial neural networks is possible by implementing learning rules such as back-propagation: the requirement for such artificial synapses is the gradual, linear and symmetric weight update upon the application of identical pulses. Further, biological learning rules such as “Spike-Timing-Dependent Plasticity” and the combination of short- and long- term potentiation or depression of the synapses paves the way to the hardware implementation of Spiking Neural Networks. A strong component of the exploratory work behind the emergence of novel devices is their fabrication. With this tutorial, we would like to give researchers and entrepreneurs an insight on how to make the first step from a concept towards a technology We will introduce the Binnig and Rohrer Nanotechnology Center (BRNC) clean room, a state-​of-the-art research facility, operated jointly by IBM Research Zurich and ETH Zurich. The laboratory offers a large number of processes and materials: we will go through a few examples of processes potentially transferable from research to industry, in particular using hafnium oxide based solid-state synapses. We will further detail why compatibility with Back-End-Of-Line integration is important and how it constrains the processing conditions.