A Hybrid Performance Prediction Approach for Fully-Connected Artificial Neural Networks on Multi-core Platforms

International audiencePredicting the performance of Artificial Neural Networks(ANNs) on embedded multi-core platforms is tedious. Concurrent ac-cesses to shared resources are hard to model due to congestion effects onthe shared communication medium, which affect the performance of theapplication. Most approaches focus therefore on evaluation through sys-tematic implementation and testing or through the building of analyticalmodels, which tend to lack of accuracy when targeting a wide range ofarchitectures of varying complexity. In this paper we present a hybridmodeling environment to enable fast yet accurate timing prediction forfully-connected ANNs deployed on multi-core platforms. The modelingflow is based on the integration of an analytical computation time modelwith a communication time model which are both calibrated throughmeasurement inside a system level simulation using SystemC. The ANNis described using the Synchronous DataFlow (SDF) Model of Computa-tion (MoC), which offers a strict separation of communications and com-putations and thus enables the building of separated computation andcommunication time models. The proposed flow enables the prediction ofthe end-to-end latency for different mappings of several fully-connectedANNs with an average of 99.5 % accuracy between the created modelsand real implementation