Hybrid Performance Prediction Models for Fully-Connected Neural Networks on MPSoC

National audience; Predicting the performance of Artificial NeuralNetworks (ANNs) on embedded multi-core platforms is tedious.Concurrent accesses to shared resources are hard to model dueto congestion effects on the shared communication medium,which affect the performance of the application. In this paperwe present a hybrid modeling environment to enable fast yetaccurate timing prediction for fully-connected ANNs deployedon multi-core platforms. The modeling flow is based on theintegration of an analytical computation time model with acommunication time model which are both calibrated throughmeasurement inside a system level simulation using SystemC. Theproposed flow enables the prediction of the end-to-end latencyfor different mappings of several fully-connected ANNs with anaverage of more than 99 % accuracy.