A Measurement-based Performance Evaluation Framework for Neural Networks on MPSoCs

National audienceEvaluation of performance for complex applicationssuch as Artificial Intelligence (AI) algorithms and morespecifically neural networks on Multi-Processor Systems on aChip (MPSoC) is tedious. Mechanisms such as data-dependentpaths and communication bus congestion induce execution timevariation, which is hard to predict accurately using traditionalanalysis methods. This paper illustrates our proposed performanceprediction workflow based on simulation models forprobabilistic timing prediction for MPSoC. We aim to extend ourexisting approach to optimize neural network implementation onresource-constrained multiprocessor platforms