The deep learning community focuses on training networks for a better accuracy on GPU servers. However, bringing this technology to consumer products requires inference adaptation of such
Instruction networks for low-energy, small-memory, and computationally constrained edge devices. ARM CPU is one of the important components of edge devices, but a clear comparison between the existing
inference frameworks is missing. We provide minimal preliminaries about ARM CPU architecture and briefly mention the difference between the existing inference frameworks to evaluate them based on performance versus usability trade-offs.