Locality aware MAC for energy-efficient CNN accelerator

Abstract

This thesis concerns designing a MAC for Convolutional Neural Network (CNN) hardware accelerator. CNN is the most promising image recognition and classification algorithm in the present and future. Many applications have already adopted CNNs for image recognitions, and more areas are expected to employ CNN in the future. The powerfulness of CNN has been proven through many researches in terms of its accuracy. However, energy efficiency should be improved for many applications to IoT and mobile devices to use CNN. Since the most workload of CNN is concentrated on the convolution operation, an energy-efficient hardware for convolution will reduce the power consumption of CNN accelerator. Therefore, I set the thesis topic as an energy efficient MAC for CNN accelerator.

I have studied CNN algorithms and design methodologies for hardware accelerators. Based on the studies, I propose a new MAC to reduce power consumption during convolution in CNN. In particular, the error-resilience and feature locality of CNN are exploited. Some multiplications are intentionally omitted to save energy, but the error caused by the skipped calculations can become negligible due to the error-resilience and feature locality of CNN. If the error resides within a scope that can be handled by CNN, the efficient power saving can be accomplished. The omitted multiplication results are shared from neighboring calculation based on data locality. Specifically, if two input data are similar, one of the two multiplications is not performed. The result of the unperformed calculation comes from the other result. That is to say, the result is shared. Post synthesis simulation of the proposed MAC shows a 20% energy saving in return for 3.6% accuracy loss.