Predicting the performance of Artificial Neural Networks (ANNs) on embedded multi-core platforms is tedious. Concurrent accesses to shared resources are hard to model due to congestion effects on the shared communication medium, which affect the performance of the application. Most approaches focus therefore on evaluation through systematic implementation and testing or through the building of analytical models, which tend to lack of accuracy when targeting a wide range of architectures of varying complexity. In this paper we present a hybrid modeling environment to enable fast yet accurate timing prediction for fully-connected ANNs deployed on multi-core platforms. The modeling flow is based on the integration of an analytical computation time model with a communication time model which are both calibrated through measurement inside a system level simulation using SystemC. The ANN is described using the Synchronous DataFlow (SDF) Model of Computation (MoC), which offers a strict separation of communications and computations and thus enables the building of separated computation and communication time models. The proposed flow enables the prediction of the end-to-end latency for different mappings of several fully-connected ANNs with an average of 99.5 % accuracy between the created models and real implementation.

When deploying Artificial Neural Networks (ANNs) onto multicore embedded platforms, an intensive evaluation flow is necessary to find implementations that optimize resource usage, timing and power. ANNs require indeed significant amounts of computational and memory resources to execute, while embedded execution platforms offer limited resources with strict power budget. Concurrent accesses from processors to shared resources on multi-core platforms can lead to bottlenecks with impact on performance and power. Existing approaches show limitations to deliver fast yet accurate evaluation ahead of ANN deployment on the targeted hardware. In this paper, we present a modeling flow for timing and power prediction in early design stage of fully-connected ANNs on multi-core platforms. Our flow offers fast yet accurate predictions with consideration of shared communication resources and scalability in regards of the number of cores used. The flow is evaluated on real measurements for 42 mappings of 3 fully-connected ANNs executed on a clock-gated multi-core platform featuring two different communication modes: polling or interrupt-based. Our modeling flow predicts timing with 97 % accuracy and power with 96 % accuracy on the tested mappings for an average simulation time of 0.23 s for 100 iterations. We then illustrate the application of our approach for efficient design space exploration of ANN implementations.