An Integer-Only Resource-Minimized RNN on FPGA for Low-Frequency Sensors in Edge-AI

Abstract

The growth of Artificial Intelligence (AI) and the Internet of Things (IoT) sensors has given rise to a synergistic paradigm known as AIoT, wherein AI functions as the decision-maker and sensors collect information. However, a substantial proportion of AIoT rely on cloud-based AI, which process wirelessly transmitted raw data, increasing power consumption and reducing battery life at sensor nodes. Edge-AI has emerged as a promising alternative, implementing AI directly on sensor nodes, eliminating the need of raw data transmission. Despite its potential, there is a scarcity of hardware architectures optimized for resource-constrained platforms, such as field programmable gate arrays (FPGAs), particularly for low-frequency sensors. This work presents a shared-scale integer-only recurrent neural network (RNN) implemented on a Lattice ICE40UP5K FPGA using a resource-minimized time and layer-multiplexed (TLM) hardware architecture. This architecture adopts real-time processing, setting clock frequency to complete a single RNN timestep preceding the next sensor sample, reducing power consumption significantly. Measurements on this FPGA implementing our proposed architecture applied to a pretrained RNN on cow behavior show a power consumption of 360 μ W at a clock frequency of 146 kHz and negligible accuracy loss at 8-bit bitwidth. This finding suggests that our methods lead to the most accurate implementation of animal behavior estimation with a power consumption below 500 μ W on an FPGA. The implementation in Systemverilog and Python code is publicly available, enabling adaptation of the RNN for various tasks involving low-frequency sensors on resource-constrained FPGAs, thereby contributing to the further advancement and democratization of Edge-AI solutions.