Edge AI must continuously adapt to dynamically changing power budgets, reliability requirements, and input distributions, which is typical in long‑running scenarios such as autonomous driving, industrial monitoring, and satellite onboard intelligence.
Existing fault‑tolerant schemes assume static operating conditions, while continual‑learning approaches ignore concurrent hardware faults, and little work has explored runtime‑adaptive reliability on programmable accelerators.
WARD introduces a runtime‑adaptive Vision Transformer framework that combines channel‑wise subnetwork partitioning, reliability‑aware continual learning, and dynamic operating‑mode scheduling. These components jointly optimize performance, fault tolerance, and adaptability at runtime.
The network is split into two physically isolated subnetworks and can execute under four modes: Full‑Precision, Low‑Power, High‑Reliability, and Adaptive. Each mode adjusts computational precision and redundancy to balance cost and reliability while guaranteeing uninterrupted real‑time inference.
To demonstrate deployability, WARD is implemented on a lightweight FPGA accelerator with added hardware support for mode scheduling and resource management. Experiments show a network‑level failure rate of only 1.79% under high bit‑error rates, less than 5% area overhead, and mode transitions completed within a few clock cycles.
These results confirm that adaptive reliability management can be integrated into programmable edge AI accelerators with negligible overhead.
Review