A Hebbian Learning based Spiking Neuromorphic Hardware Accelerator with ReRAM

Green Mountain Semiconductor (GMS) is presenting functional subcomponents designed for the core of a spiking neural network (SNN). This design utilizes Resistive Random-Access Memory (ReRAM) as a radiation tolerant storage element. SNNs are a biologically inspired computing paradigm that more closely resemble neural processing in the human brain than conventional artificial neural networks. SNNs communicate through discrete pulses, with information encoded in their timing or rate. This event-driven nature allows these networks to achieve greater energy efficiency than the state of the art. By leveraging ReRAMs ability to efficiently store and update synaptic weights, the proposed system enables energy-efficient, scalable, and resilient artificial intelligence processing. This architecture supports unsupervised Hebbian learning SNN models, mimicking biological learning mechanisms to achieve in situ adaptability in real time. A proof of concept circuit was implemented with a two neuron input layer and a two neuron output layer, fully connected through a resistive memory array having 2 bit lines and two word lines. The neurons are designed as fully asynchronous leaky integrate and fire circuits without any clock signal. The design has recently been submitted for manufacturing at GlobalFoundries in 22nm FDX technology. Scaled power consumption and compute performance will be discussed and benchmarked against state of the art solutions, along with further improvements to reduce power consumption for a variety of technology nodes.