Early Timing and Energy Prediction and Optimization of Artificial Neural Networks on Multi-Core Platforms

The need to implement artificial Neural Networks (NNs) on embedded multicore platforms has become fundamental. Predicting timing properties (inference time, latency, throughput) and energy as early as possible in the design process is necessary to find solutions that optimize resource use and respect the constraints imposed on the system. A major modeling difficulty comes from the need to correctly describe the influence of resource sharing (processor, memory, communication bus) within multi-core platforms. In this thesis, we present a complete flow for predicting and optimizing timing properties and energy, combining several modeling approaches. This flow leads to optimized resource occupancy without degrading the performance of implemented NNs. Predictions are compared with measurements on real targets. The proposed models have an accuracy of over 97% on timing and 93% on energy for 54 mappings of 4 NNs, with a prediction time of 20s per mapping. We show how to use the models to efficiently explore the design space and find optimized solutions that satisfy the constraints imposed on the system.