Project Details
Projekt
Modeling of near-memory computing architectures for computer vision applications targeting performance and energy
Applicant
Professorin Dr.-Ing. Diana Goehringer
Subject Area
Computer Architecture, Embedded and Massively Parallel Systems
Term
since 2026
Project identifier
Deutsche Forschungsgemeinschaft (DFG) - Project number 545776403
A large fraction of the execution time and the energy costs of modern data-intensive workloads are spent by moving data between memory units and processing cores. New computing paradigms such as near- and in-memory computing reduce the energy needed for data transport by having processing units next to or even in the memory units. This is an important step to enable a sustainable computing solution as targeted by the research unit Mod4Comp. The project within Mod4Comp focuses on a novel approach for modeling, simulation and hardware generation of near-memory computing architectures (NMAs). Machine learning (ML) assisted algorithms will be investigated and realized to identify compute kernels for NMAs and their best location within the memory hierarchy of the processor (L1-, L2, L3-cache, main memory) in order to optimize the energy-efficiency and the computational performance of the overall applications. A library of computer vision (CV) and embedded artificial intelligence (AI) proxy apps will be generated for evaluation. Performance and energy models for NMAs will be designed and validated using simulation as well as an FPGA prototype. The results, i.e. the CV and embedded AI proxy apps, the ML-assisted algorithms for identification and localization of NMAs within the memory hierarchy of the processor, the hardware generation framework for NMAs and the performance and energy models will be integrated in close collaboration with the other subprojects of the research unit into the holistic Mod4Comp multi-layer modeling workflow.
DFG Programme
Research Units
