A 28nm Energy-efficient Asynchronous SNN Accelerator with On-chip Learning for Gas Recognition

In this paper, we present an energy-efficient asynchronous spiking neural network (SNN) accelerator enabling on-chip learning for gas recognition, which contains 144 neurons and 4.8K synapses with 4-bit weight. Instead of a synchronous high-frequency clock, based on the two-phase handshake protocol, we adopt Click-based bundled-data asynchronous circuits to implement the SNN algorithm with spike timing-dependent plasticity (STDP) with lateral inhibition. The SNN accelerator is fabricated by TSMC 28nm CMOS process, and the post place and router (post-PAR) simulation results show that based on the publicly available UCI gas dataset, our accelerator can recognize 10 types of gas under 40% noise with one-shot learning, achieving 90.9% accuracy and 2.38SpJ/SOP energy efficiency at 0. 9V. Thus, we believe that as a gas recognition circuit with on-chip learning, it provides a promising solution for the low-power portable intelligent E-nose system.