2

Estimating the energy efficiency of DNNs design is challenging since it largely depends on the data movement and we demonstrated that the conventional arithmetic intensity is a poor proxy for the energy efficiency in DNNs. We proposed a novel metric termed “Data-type aware arithmetic intensity” (DI) which gives unequal importance to weight and activation reuse and more accurately estimates the energy efficiency of a wide range of DNNs.

The remarkable predictive performance of deep neural networks (DNNs) has led to their adoption in service domains of unprecedented …