Move the Roof: Model-driven Methodology for Designing Efficient Deep Learning Architectures
Paper i proceeding, 2026
Designing efficient architectures for Deep Learning (DL) at the edge is challenging, as the growing demands of emerging models contrast with the tight resource constraints. This work proposes a model-driven methodology for the fast exploration of large design spaces, producing efficient architectures under performance constraints. Our methodology combines a strategy to aggressively prune the design space with a Cache-Aware Roofline Model (CARM)-based performance estimation to avoid costly simulations.
runtime
efficiency
edge
design space exploration
deep learning
design methodology
Författare
Alexandre Rodrigues
Instituto de Engenharia de Sistemas e Computadores: Investigação e Desenvolvimento em Lisboa
Mateo Vázquez Maceiras
Instituto de Engenharia de Sistemas e Computadores: Investigação e Desenvolvimento em Lisboa
Mohammad Ali Maleki
Instituto de Engenharia de Sistemas e Computadores: Investigação e Desenvolvimento em Lisboa
Muhammad Waqar Azhar
Chalmers, Data- och informationsteknik, Datorteknik
Göteborgs universitet
Aleksandar Ilic
Universidade de Lisboa
Pedro Petersen Moura Trancoso
Chalmers, Data- och informationsteknik, Datorteknik
Göteborgs universitet
Leonel Sousa
ZeroPoint Technologies AB
Proceedings of the ACM Symposium on Applied Computing
664-666
9798400722943 (ISBN)
41st Annual ACM Symposium on Applied Computing, SAC 2026
Thessaloniki, Greece,
Thessaloniki, Greece,
The European Processor Initiative (EPI)
Europeiska kommissionen (EU) (EC/H2020/800928), 2018-12-01 -- 2021-11-30.
European, extendable, energy-efficient, energetic, embedded, extensible, Processor Ecosystem (eProcessor)
Europeiska kommissionen (EU) (EC/H2020/956702), 2021-01-01 -- 2024-06-30.
EPI SGA2
Europeiska kommissionen (EU) (101036168), 2022-01-01 -- 2024-12-31.
Ämneskategorier (SSIF 2025)
Datavetenskap (datalogi)
DOI
10.1145/3748522.3779938