Empirical Power-performance Analysis of Layer-wise CNN Inference on Single Board Computers

This paper analyzes the impact of input sparsity and DFS/DVFS configurations for single-board computers on execution time, power, energy each VGG16 layer as first step towards efficient CNN inference computers. For this purpose, we develop a power time measurement environment perform experiments using Raspberry Pi 4 NVIDIA Jetson Nano. Our results show that clock frequency strongly correlates with power. Inversely, has weak correlation Then, coarse-grained DVFS model can explain over 96% variations in even when sets voltage computer are unavailable.