A CNN-inspired mixed signal processor based on tunnel transistors

Novel devices are under investigation to extend the performance scaling trends that have long been associated with Moore´s Law-based device scaling. Among the emerging devices being studied, tunnel FETs (or TFETs) are particularly attractive, especially when targeting low power systems. This paper studies the potential of analog/mixed-signal information processing using TFETs. The design of a highly-parallel processor - inspired by cellular neural networks - is presented. Signal processing is performed partially in the time-domain to better leverage the unique properties of TFETs, i.e., (i) steep slopes (high g_m/I_DS) in the subthreshold region, and (ii) high output resistance in the saturation region. Assuming an InAs TFET with feature sizes comparable to the 14 nm technology node, a power efficiency of 10,000 GOPS/W is projected. By comparison, state-of-the-art hardware assuming CMOS technology promises a power efficiency only close to 1,000 GOPS/W.