MEDEA: A Design-Time Multi-Objective Manager for Energy-Efficient DNN Inference on Heterogeneous Ultra-Low Power Platforms

The growing demand for on-device AI necessitates energy-efficient execution of DNN based applications on resource-constrained ultra-low power (ULP) platforms. Heterogeneous architectures, combining specialized processing elements (PEs), have emerged as a key solution for achieving the required performance and energy efficiency. However, optimizing energy while executing applications on these platforms requires efficiently managing platform resources like PEs, power features, and memory footprint, all while adhering to critical application deadlines. This paper presents MEDEA, a novel design-time multi-objective manager for energy-efficient DNN inference on Heterogeneous ULP (HULP) platforms. MEDEA uniquely integrates: kernel-level dynamic voltage and frequency scaling (DVFS) for dynamic energy adaptation; kernel-level granularity scheduling, suitable for specialized accelerators; memory-aware adaptive tiling to navigate severe memory constraints; and all within a timing constraint-based optimization strategy, which minimizes energy based on application deadline. To showcase practical viability, we evaluate MEDEA on HEEPtimize, a heterogeneous ULP platform (22 nm, FPGA-prototyped) featuring a RISC-V processor besides Near-Memory Computing (NMC) and Coarse-Grained Reconfigurable Array (CGRA) accelerators. Experimental results, using a biomedical seizure detection case study, demonstrate that MEDEA achieves overall energy reductions of up to 38% compared to representative state-of-the-art methods, while consistently meeting all timing and memory requirements. This effectiveness is attributed to its integrated features, with our analysis showing that kernel-level DVFS alone can be responsible for over 31% of the energy savings in specific scenarios.