Lightweight Hardware/Software Co-Design Framework for Real-Time Reconfigurable Accelerators in IoT Devices

Abstract

The paper proposes a lightweight hardware/software co-design framework, which is proposed to support real-time reconfigurable acceleration in resource limited Internet of Things (IoT) devices. As demands increase at the network edge, it is becoming stressed by static, hardware architectures as more compute is being asked to be pushed. To overcome these difficulties the proposed framework schemes partial reconfiguration, dynamic task offloading, and runtime scheduling as a combination to maximize performance with severe resource budget constraints. The system takes a dynamic approach of redistributing tasks to the software and hardware space by using a minimal decision engine and modular hardware accelerators to achieve adaptiveness in response to the workload nature. Application and testing against a Xilinx Artix-7 and a Lattice iCE40 platform generates a large decrease in the computational latency, of around 36% and energy utilization of 42%, as compared to the traditional arrangement of static accelerator platforms. These findings confirm the possibilities of the framework to be scalable, flexible, and energy saving in real time to IoTs applications like sensor analytics, environmental monitoring and intelligent automations.