A Compact and All-Round Spiking Artificial Neuron with Diffusive Memristor for Neuromorphic Computation

Artificial intelligence (AI) is approaching a stage where the limitations of current hardware technology in terms of energy and area efficiency significantly hinder system performance. Despite the deployment of tens of thousands of GPUs and megawatt-level power consumption, the most advanced AI system still doesn’t match the intelligence of the human brain, which operates at only 20W within a volume of one cubic decimeter. To overcome this challenge, neuromorphic computing aims to leverage emerging devices and circuits to emulate the behavior of biological neural networks for more powerful AI hardware. Here, we have developed a highly scalable and energy-efficient spiking artificial neuron, composing one asymmetric diffusive memristor (ADM), one transistor and one resistor, referred to as the 1M1T1R neuron.<br/><br/>The ADM is a type of memristive devices with asymmetric switching behavior due to the migration of conductive ions within the asymmetric device. We compared two types of ADMs with HfO₂ and SiO₂ switching layers and observed different operation voltages and currents. The HfO₂ ADM was selected for constructing the 1M1T1R neuron due to its suitable operation range. The rich and asymmetric ion dynamics of the ADM, combined with the unique circuit configuration of the 1M1T1R neuron, enabled us to experimentally demonstrate six critical neuronal functions: leaky integration, threshold firing, cascaded connection, intrinsic plasticity, refractory period, and stochasticity. These functions cover all essential behaviors for general spiking neural network applications. The 1M1T1R neuron demonstrated an energy consumption of just 1pJ per spike and has the potential to be scaled down to the size of a single transistor by stacking the memristor and resistor on the transistor's gate.<br/><br/>We also developed a compact model of the 1M1T1R neuron that faithfully reproduces its functionality. Utilizing this model, we designed a recurrent spiking neural network (RSNN) with customized forward and backward algorithms based on the 1M1T1R neuron's behaviors. The RSNN was tested on the Spiking Heidelberg Digits (SHD) spatiotemporal dataset, achieving an average testing accuracy of 91.35%. Additionally, tuning the neuron functions during RSNN testing resulted in more than a 10% variation in network performance, highlighting the significance of these functions. These findings pave the way towards highly energy- and area-efficient neuromorphic systems by exploiting the ion dynamics in emerging memristive devices and co-designing hardware behaviors and software algorithms.