Stochastic neuron design using conductive bridge RAM

摘要

We present an original methodology to design hybrid neuron circuits (CMOS + non volatile resistive memory) with stochastic firing behaviour. In order to implement stochastic firing, we exploit unavoidable intrinsic variability occurring in emerging non-volatile resistive memory technologies. In particular, we use the variability on the ‘time-to-set’ (tset) and ‘off-state resistance’ (ROff) of Ag/GeS2 based Conductive Bridge (CBRAM) memory devices. We propose a circuit and a novel self-programming technique for using CBRAM devices inside standard Integrate and Fire neurons. Our proposed solution is extremely compact with an additional area overhead of 1R-3T. The additional energy consumption to implement stochasticity in Integrate and Fire neurons is dominated by the CBRAM set-process. These results highlight the benefits of novel non memory technologies, whose impact may go far beyond traditional memory markets.