Performance Analysis of Convolutional Neural Networks on Embedded Systems

Machine learning is no longer confined to cloud and high-end server systems and has been successfully deployed on devices that are part of Internet of Things. This paper presents the analysis of performance of convolutional neural networks deployed on an ARM microcontroller. Inference time is measured for different core frequencies, with and without DSP instructions and disabled access to cache. Networks use both real-valued and complex-valued tensors and are tested using different inference engines. We conclude that the system must be tuned in a holistic way to achieve optimal efficiency.