Energy Efficient and High-Performance Synaptic Operating Point Evaluation for SNN Applications

The synapses are fundamental components of spiking neural networks (SNNs). Although metal-oxide memristors, including Hafnium-Oxide (Hf0
2
), offer potential in synaptic circuits, the design process must account for their non-idealities. This paper proposes a current-controlled memristive synapse with a limited range of operation close to the low resistance state (LRS) to address the device non-idealities experienced at the high resistance state (HRS) and enhance reliability. The operating voltage of the synapse is optimized to maximize performance considering process variation, current resolution, and SNN performance. The impact of adjusting the operating voltage is evaluated with three datasets. At last, application-specific operating points are determined within a narrow range that balances current resolution, energy consumption for SNN applications, and reliability.