Reconfigurable Computing and FPGAs for Adaptive AI Workloads

Reconfigurable Computing and FPGAs for Adaptive AI Workloads

Anant Manish Singh

anantsingh1302@gmail.com

Department of Computer Engineering

Thakur College of Engineering and Technology, Mumbai, Maharashtra, India

Divyanshu Brijendra Singh

singhdivyanshu7869@gmail.com

Department of Computer Engineering

Thakur College of Engineering and Technology (TCET), Mumbai, Maharashtra, India

Aditya Ratnesh Pandey

ap7302758@gmail.com

Department of Computer Engineering

Thakur College of Engineering and Technology (TCET), Mumbai, Maharashtra, India

Maroof Rehan Siddiqui

maroof.siddiqui55@gmail.com

Department of Computer Engineering

Thakur College of Engineering and Technology (TCET), Mumbai, Maharashtra, India

Akash Pradeep Sharma

sharmaakash22803@gmail.com

Department of Computer Engineering

Thakur College of Engineering and Technology (TCET), Mumbai, Maharashtra, India

Amaan Zubair Khan

hhkhananamaan@gmail.com

Department of Computer Engineering

Thakur College of Engineering and Technology (TCET), Mumbai, Maharashtra, India

Abstract

Reconfigurable computing using Field-Programmable Gate Arrays (FPGAs) is rapidly emerging as a critical solution for adaptive and efficient AI processing. The ability to dynamically adapt hardware resources to changing AI workloads enables significant gains in energy efficiency, latency and throughput, especially for edge and real-time applications. In this work, we present a novel Dynamically Reconfigurable AI Processor (DRAP) framework that leverages dynamic partial reconfiguration (DPR), hardware-software co-design and multicast-optimized dataflow to address major bottlenecks identified in recent literature. Real-time experiments on Xilinx Virtex-7 and Intel Stratix 10 NX platforms demonstrate up to 12.6× higher TOPS/W efficiency and 38% lower inference latency compared to leading GPU solutions. Our approach achieves 89% logic utilization and a 53.8% reduction in energy consumption for pruned convolutional neural networks (CNNs) in video analytics. We validate our results with industry-relevant workloads and compare DRAP to state-of-the-art methods, showing that our framework mitigates key gaps such as static resource allocation, inefficient multicast handling and thermal bottlenecks. This paper provides a comprehensive analysis of dynamic hardware adaptation strategies for FPGAs offering actionable insights for deploying adaptive AI at scale.