Tiny Neuron Network System based on RISC-V Processor: A Decentralized Approach for IoT Applications

Abstract

The idea of Artificial Intelligence of Things (AIoT), a combination of Artificial Intelligence, especially Deep Learning, with edge devices in IoT networks, has recently emerged to reduce the communication cost, and server workloads and improve user experiences. This work presents our current research on a tiny neural network accelerator in a RISC-V System-on-Chip (SoC) to accelerate AI in IoT applications. This accelerator implements a variable-bit-precision MAC or a stochastic MAC to reduce hardware area and power consumption. The tiny AI accelerator has been successfully integrated into a low-power IoT SoC. The design has been implemented on FPGA technology using Arty A7 100T development kit with the operating frequency of 50MHz and the hardware resource of 12K slices. For the MNIST dataset, the accelerator with 8-bit precision can perform Convolutional Neural Network with an accuracy of 98.55%.