MAGNETO: A Genetic Algorithm-Based Power-Aware Mapping Optimization Framework for Mobile NPUs

MAGNETO is a power-Aware genetic search framework for mapping deep neural network layers onto neural processing units with strict power constraints. MAGNETO incorporates a penalty-based constraint enforcement mechanism and an energy-delay product-oriented fitness function to guide the mapping search toward solutions that balance latency and energy efficiency. Unlike traditional mapping strategies, MAGNETO dynamically explores mapping configurations to discover high-quality mappings tailored to each layer's computational and memory access characteristics.Extensive experiments across various layer types-including convolutional layers, general matrix multiplication layers, and linear layers-demonstrate that MAGNETO consistently outperforms baseline strategies in terms of TOPS/W and Energy per MAC. Notably, MAGNETO also achieves both low latency and efficient energy usage even under a strict 1 W power budget, showing competitive or superior latency-energy trade-offs. Our results highlight the potential of search-based mapping under power constraints for real-Time, energy-efficient inference on NPUs in edge environments.