Performance Evaluation of Systolic DCNN Accelerators

Abstract

For the given deep convolutional neural network (DCNN) models, in this work, we carefully evaluate systolic DNN accelerators associated with different array configurations. In order to find the optimal array structure in terms of the processing efficiency, the open-source evaluation framework is adopted for investigating the relationships between the network architecture and the hardware-level options. We also present the selection criteria of array structure that leads to the efficient accelerator designs.