Work

A Dynamically Reconfigurable FPGA Architecture for Energy-Efficient Embedded Intelligence

Year: 2025
Type: article
Abstract: The basic design requirement of embedded intelligence systems at the network end is energy efficiency, as limited power supply has to be balanced between power scarcity requirements and strict real time performance specifications. Traditional general-purpose computers are not energy efficient when used in long-term intelligent applications, and statically programmed hardware accelerators are inefficient at resource over-allocation and are ineffective in adapting to the dynamic workload variations. This paper will suggest a dynamically reconfigurable FPGA architecture to overcome these limitations so as to have energy-proportional embedded intelligence due to the possibilities it provides in the administration of the computational resources at runtime. The architecture divides up the FPGA into a fixed stable partition and a range of modular reconfigurable ones so that application-specific accelerators can be deployed, resized or switched off using fine-grained partial reconfiguration. The workload-conscious runtime manager is a lightweight runtime manager that measures the execution requirements and executes resilient reconfiguration choices that never interfere with the systems execution. Coordinated power gating and configuration-level optimization is also used to achieve energy efficiency, so that the active hardware is immediately adjusted to the current workload requirements. The design methodology suggested allows supporting many performance/energy trade-offs in different variants of accelerators and allows to operate efficiently with a wide range of embedded AI workloads. Experimental analysis of representative applications of edge-intelligence by experimental evaluation shows that, aside from introducing predictable latency and real-time behavior, the proposed architecture draws significantly lower levels of overall energy consumption than state-of-the-art FPGA architecture designs, as well as compared to processor-based implementation applications. The results of this work indicate that dynamic hardware reconfiguration is an effective architectural concept in the next-generation energy-efficient embedded intelligence systems.
Language: English
Open Access status: diamond
OpenAlex ID: W7219948550
FWCI: 0.366
Cites: 4
Cited by: 1
Related to: 10