Evaluation of Task Mapping on Multicore Neural Network Accelerators


Satoshi SHINDO, Momoka OHBA, Tomoaki TSUMURA, Shinobu MIWA : "Evaluation of Task Mapping on Multicore Neural Network Accelerators", Proc. 4th Int'l Workshop on Computer Systems and Architectures (CSA'16) ,pp415--421 (Nov. 2016) Proceeding


Deep neural networks are widely used for many applications such as image classification, speech recognition and natural language processing because of the high recognition rate. Since general-purpose processors such as CPUs and GPUs are not energy efficient for such neural networks, application-specific hardware accelerators for neuralnetworks (a.k.a. neuralnetwork accelerators) have been proposed to improve the energy efficiency. There are many studies to increase the energy efficiency of neural network accelerators, but few studies focus on task allocation on the accelerators. This paper provides the first exploration of task mapping to cores within neural network accelerators for the increased performance. Intuitively, a well-tuned task mapping has less amount of communication between cores. To confirm this assumption, we tested two types of task mappings that generate different amount of communication between cores on a neural network accelerator. Our experimental results show that the number of communication between cores strongly affects the execution cycle of the neural network accelerator and the most effective task mapping differs depending on the size of neural networks.