Abstract

Edge AI accelerators have emerged as a critical component for real time inference under strict power and latency constraints. Conventional accelerator architectures have focused on exact computation, which has limited the achievable energy efficiency when deployed in resource-constrained edge environments. Approximate computing has gained attention as a promising paradigm that has traded controlled accuracy loss for significant gains in power and performance. However, most existing approximate designs have remained static and application-specific, which has reduced their adaptability across diverse AI workloads. The primary challenge has involved designing an architecture that has supported dynamic accuracy–energy trade-offs while maintaining acceptable inference quality. Fixed approximation levels have failed to respond to varying workload sensitivities, data distributions, and quality-of-service requirements. As a result, edge AI systems have suffered from either unnecessary energy consumption or unacceptable accuracy degradation. This work has proposed a reconfigurable approximate computing architecture that has enabled runtime adaptation of approximation levels within an edge AI accelerator. The architecture has integrated configurable approximate arithmetic units, adaptive precision control, and a lightweight reconfiguration controller that has monitored workload characteristics. Approximation modes that have targeted multipliers, adders, and accumulation paths have been selectively activated based on layer-wise sensitivity analysis. A design framework that has supported rapid switching between accuracy modes has been implemented and evaluated using representative convolutional and transformer-based inference workloads. Experimental evaluation demonstrates that the proposed architecture reduces energy consumption from 3.3 mJ to 3.05 mJ across thresholds (?_1=0.1 to ?_3=0.3) while maintaining inference accuracy within 1.9% deviation of the exact baseline. Compared with the exact baseline accelerator, energy savings reach up to 36%, and latency decreases from 16.2 ms to 15.4 ms. Energy–accuracy efficiency (?) achieves 0.75, outperforming static and learning-based approximate accelerators. These results indicate that sensitivity-aware reconfigurable approximation effectively balances energy efficiency and output quality, providing a practical solution for diverse edge AI workloads.

Authors

Pitty Nagarjuna¹, Shaik Mohammed Rizwan²
Indian Institute of Science, Bengaluru, India¹, Jazan University, Soudi Arabia²

Keywords

Approximate Computing, Edge AI Accelerators, Reconfigurable Architecture, Energy Efficiency, Adaptive Precision