Early Performance and Energy Prediction of Neural Networks Deployed on Multi-Core Platforms
Quentin Dariol · HAL (Le Centre pour la Communication Scientifique Directe) · 2023
Early evaluation of Neural Networks (NN) deployments on multi-core platforms is necessary to find deployments that optimize resource usage, performance and energy. In this paper, we propose a timing and power modeling methodology which combines simulation, analytical models, and measurements to offer fast yet accurate performance and energy prediction of NN deployments on multi-core platforms. The proposed approach is validated against measurements obtained from a real implementation of 27 mappings of four NNs with high accuracy and a fast evaluation time of approximatively 20 s per mapping.