Power management is a major challenge for embedded AI systems at the network edge. These systems must run machine learning workloads under tight energy limits. An effective solution needs a multi-layered approach. Key elements include the Power State Coordination Interface (PSCI), secure firmware, and Linux kernel features for runtime control. Core techniques such as dynamic frequency scaling, clock gating, suspend/resume, and memory or accelerator-specific optimizations further improve efficiency. Environmental factors add to the challenge. Automotive and industrial systems must meet strict thermal limits. Battery-powered devices face even tighter energy budgets. Both require adaptive control strategies. From a software perspective, effective methods include specialized kernel drivers, standardized power APIs, and optimizations such as dynamic logic gating and real-time power monitoring. When combined, these enable power-efficient AI systems that maintain reliable performance while staying within thermal and energy boundaries.
Senthil Nathan Thangaraj (Sat,) studied this question.