Modeling Data Reuse in Deep Neural Networks by Taking Data-Types into Cognizance

In recent years, researchers have focused on reducing the model size and number of computations (measured as “multiply-accumulate” or MAC operations) DNNs. The energy consumption a DNN depends both operations efficiency each operation. former can be estimated at design time; however, latter intricate data reuse patterns underlying hardware architecture. Hence, estimating it time is challenging. This article shows that conventional approach to estimate reuse, viz. arithmetic intensity, does not always correctly degree in DNNs since gives equal importance all types. We propose novel model, termed “data type aware weighted intensity” (DI), which accounts for unequal different types evaluate our 25 state-of-the-art two GPUs. show accurately models data-reuse possible convolution layers. better indicator also its generality using central limit theorem.