Work-in-Progress: Ultra-fast yet Accurate Performance Prediction for Deep Neural Network Accelerators

Abstract

We present an automatic methodology to accurately predict the performance of Deep Neural Network (DNN) accelerators using abstract descriptions of accelerator architectures and DNNs with a high degree of flexibility. By mapping partially unrolled neural network layers onto accelerator architectures, we automatically construct an analytical performance model, exploiting the dataflow-driven nature of DNNs that allows us to evaluate only a few loop iterations to determine the performance of a whole DNN layer.