Abstract
Multi-layer Perceptron (MLP) is a class of Artificial Neural Networks widely used in regression, classification, and prediction. To accelerate the training of MLP, more cores can be used for parallel computing on many-core systems. With the increasing number of cores, interconnection of cores has a pivotal role in accelerating MLP training. Currently, the chip-scale interconnection can either use electrical signals or optical signals for data transmission among cores. The former one is known as Electrical Network-on-Chip (ENoC) and the latter one is known as Optical Network-on-Chip (ONoC). Due to the differences of optical and electrical characteristics, the performance and energy consumption of MLP training on ONoC and ENoC can be very different. Therefore, comparing the performance and energy consumption between ENoC and ONoC for MLP training is worthy of study. In this paper, we first compare the differences between ONoC and ENoC based on a parallel MLP training method. Then, we formulate their performance model by analyzing communication and computation time. Furthermore, the energy model is formulated according to their static energy and dynamic energy consumption. Finally, we conduct extensive simulations to compare the performance and energy consumption between ONoC and ENoC. Results show that compared with ENoC, the MLP training time of ONoC is reduced by 70.12% on average and the energy consumption of ONoC is reduced by 48.36% under batch size 32. However, with a small number of cores in MLP training, ENoC consumes less energy than ONoC.