In the scope of an ongoing collaboration between the IETR lab and the German Aerospace Center (DLR), we are looking for a PhD candidate on the topic of "System Level Modeling for Timing and Energy Prediction of Artificial Neural Networks on Heterogeneous Multi-Core Platforms with Accelerators".
This three year project will be led 18 months at the IETR lab in Nantes, France and 18 months at DLR in Oldenburg, Germany.
Context
The market for Artificial Intelligence (AI) applications, particularly Artificial Neural Networks (NNs), has experienced tremendous growth in the past decade. Executing NNs on embedded multi-core platforms presents a significant challenge due to resource scarcity, real-time constraints, and energy budgets. Optimizing memory and computational resource usage is crucial to meet application constraints. To enable the efficient execution of AI applications, modern platforms often include hardware accelerators and use Instruction Set Architectures (ISA) that can be substantially upgraded to offer optimizations for the processing of AI. In the implementation of NNs, joint optimization between the multi-core system and AI accelerator is necessary to enhance latency and energy. Using models to predict non functional properties such as timing and energy can significantly reduce search time and ensure the rapid identification of optimized NN implementations. Reliably finding in affordable time embedded AI implementations that meet the strict requirements is key to enable new applications in the transportation, avionics and space domains, in which DLR is a major actor.
Proposition
In this context, we aim at proposing a hybrid modeling approach to offer fast yet accurate evaluation of NN implementation on heterogeneous multi-core platforms with possible ISA extensions and HW accelerator. The envisioned flow should take as inputs NN hyper-parameters, the specified multi-core platform, and user-defined constraints (e.g. maximum energy budget). The use of system-level simulation allows obtaining estimates with accurate modeling of shared resource effects. Timing, power, and memory use models are characterized through measurements performed on a real platform, providing highly accurate estimations. Due to possible variations in timing, probabilistic modeling will be considered to appropriately capture and provide confident prediction of external memory and caches effects.
Expected skills of the candidate
Contacts
More details available on the website of the IETR laboratory:
https://www.ietr.fr/sites/www.ietr.fr/files/medias/files/System%20Level%20Modeling%20for%20Timing%20and%20Energy%20Prediction.pdf
(c) GdR IASIS - CNRS - 2024.