AI-Augmented Runtime Reconfiguration for Energy-Aware FPGA-Based Edge Computing Systems

Abstract

In era of ubiquitous edge intelligence, Field-Programmable Gate Arrays (FPGA) have become strong platform, when it comes to performing high-performance computing with little energy consumption on resource-constraint applications. But more classical static FPGA configurations are not necessarily flexible enough to respond to variable workloads and energy requirements such as those found in edge applications. As it is explained in this paper, the AI-augmented system is a runtime reconfiguration framework compatible with FPGAs-based edge systems and continually intelligently manages hardware resources. The framework provides the ability to make in realtime decision associated with partial reconfiguration (PR) via integrating lightweight machine learning (ML) model into the system control logic code in order to optimize the task placement on hardware by taking into consideration the workload patterns, the thermal profiles and the power consumption trend. The architecture takes advantage of modular reconfigurable regions (RRs) in the FPGA fabric enabling dynamic; accelerators swapping without shutting down the system. An AI scheduler with low overheads observes the metrics of the system and forecasts on possible optimal reconfiguration steps to ease the trade-off between performance and energy efficiency. The suggested approach is applied and tested in a Xilinx Zynq system following a set of various heterogeneous edge workloads, such as convolutional neural networks (CNNs), signal processing applications, and data analytics kernels. Energy saving of up to 42 percent, throughput gain of 31 percent and 50 percent average through reconfiguration latency reduction over the baseline requirements in the static and rule-based systems have been experimentally observed. The system was also highly adaptable to real-time changes in workloads which demonstrates the scalability of AI-augmented runtime reconfiguration as a solution to next-generation edge computing infrastructures.