AERO: A 1.28 MOP/s/LUT Reconfigurable Inference Processor for Recurrent Neural Networks in a Resource-Limited FPGA

This study presents a resource-efficient reconfigurable inference processor for recurrent neural networks (RNN), named AERO. AERO is programmable to perform on RNN models of various types. was designed based the instruction-set architecture specializing in processing primitive vector operations that compose dataflows models. A versatile vector-processing unit (VPU) incorporated every operation and achieve high resource efficiency. Aiming at low usage, multiplication VPU carried out basis an approximation scheme. In addition, activation functions are realized with reduced tables. We developed prototype system using resource-limited field-programmable gate array, under which functionality verified extensively tasks several different The efficiency found be as 1.28 MOP/s/LUT, 1.3-times higher than previous state-of-the-art result.