Seq2seq Architecture with Memory-Centric Neural Network Processor towards Edge AI Computing

Abstract

The rapid advancement of generative AI has been driven by the introduction of the Transformer model [1] [2]. However, a significant challenge with such models lies in their high computational requirements. As a result, generative AI predominantly relies on online services. Nevertheless, exploring generative AI capabilities on the edge holds significant potential for enabling novel applications. By employing lightweight models, it becomes feasible to design architectures specifically for edge applications.