Séminaire Cambium, Inria Paris
                                 Grace Hopper
                         Vendredi 29 novembre, 10h30

                                 Chiara Daini
                                 Inria Paris

        Optimizing CNN Inference on Multicore Scratchpad Architectures

Convolutional neural networks (CNNs) are widely recognized for their
suitability in image recognition tasks within constrained embedded
systems, as they offer an effective trade-off between computational
complexity and memory demands. This research focuses on the challenge of
implementing CNN inference efficiently on a multicore computing platform
equipped with scratchpad memories, commonly used in high-performance
embedded applications.

To address this, an Integer Linear Programming (ILP) model is introduced,
which accounts for the costs associated with data transfer to and from
scratchpad memory, as well as the parallel processing workloads
distributed across multiple cores. This model is designed to meet strict
real-time temporal requirements while optimizing system resources.

The ILP approach includes two distinct optimization techniques that
provide flexibility between shorter analysis times and higher processing
performance, catering to various application needs. Benchmark evaluations
using standard datasets illustrate that this ILP-based solution enhances
the efficiency of CNNs, supporting more practical deployment in
real-world, time-sensitive scenarios.

            Vous pouvez vous abonner à nos annonces de séminaires:
                     http://cambium.inria.fr/seminar.html

             Nos séminaires sont accessibles en ligne en direct
                            via le lien ci-dessus.