Memory-Centric Computing for Image Classification Using SNN with RRAM

Neuromorphic computing, exemplified by spiking neural networks (SNN), seeks to replicate human brain functionality through event-driven processes, encoding information via spikes, and adopting biological learning principles. Its comparative advantage over traditional computing lies in the event-driven nature of computations, promising notably high energy efficiency. However, the hardware implementation of SNN poses limitations for various applications. This study proposes an In-memory Computing (IMC) approach, utilizing a Resistive RAM-based (RRAM) crossbar array to expedite the SNN algorithm. The investigation scrutinizes the accuracy of three network variants - fp32, fp16, and int8 - utilizing different data types. Remarkably, by reducing the datasize to one fourth of the original size, the accuracy increased by 1.17% after retraining. Additionally, quantizing the network from fp32 to 8-bit fixed point, and using an RRAM crossbar array, yielded savings of ~1634x in memory access energy, ~1636x in memory access latency, and ∼132x in computations energy. Furthermore, utilizing the RRAM crossbar array for the acceleration of the quantized SNN algorithm yielded ∼10x reduction in average power consumption per inference, and ~159x savings in required area.