Top gate overlaid carbon nanotube transistor electronic synapse arrays for high-performance image recognition

Abstract

Carbon nanotube field-effect transistor (CNTFET) based electronic synapses have great potential for brain-like neuromorphic computing, due to their low power consumption. However, the realization of diverse biological synaptic plasticity in the CNTFET remains a significant challenge due to its small dynamic range, abrupt conductance modulation and limited hardware structure. In this work, we developed a top gate overlaid carbon nanotube field effect transistor (TGO-CNTFET) with a large dynamic range, which successfully mimics synaptic functions, including excitatory and inhibitory synaptic behaviors (EPSC/IPSC), paired-pulse facilitation and depression (PPF/PPD), and spike-timing-dependent plasticity (STDP). We further investigated the synaptic performances of as-fabricated and air-annealed device arrays. Compared with as-fabricated devices, the air-annealed TGO-CNTFET array demonstrated better performance in terms of the dynamic range of STDP and the power consumption, with the latter achieving a power consumption per spike of 1.27 pJ. This improvement is further reflected in the image recognition task on the CIFAR-100 database using ResNet 50, where the air-annealed device achieved an accuracy of 93.2%, whereas the as-fabricated counterpart reached only 90.8%. This work introduces an architectural strategy for developing neuromorphic computing systems that incorporate functional oxides as dielectric layers in TGO-CNTFET-based synapses.