Project Details
Projekt
Energy and performance modeling of the neuromorphic supercomputer SpiNNcloud Subproject SP6 within the proposed Research Unit "Holistic Energy and Performance Modeling for Sustainable Computing (Mod4Comp)"
Subject Area
Computer Architecture, Embedded and Massively Parallel Systems
Hardware Systems and Architectures for Information Technology and Artificial Intelligence, Quantum Engineering Systems
Hardware Systems and Architectures for Information Technology and Artificial Intelligence, Quantum Engineering Systems
Term
since 2026
Project identifier
Deutsche Forschungsgemeinschaft (DFG) - Project number 545776403
The computing demand for artificial intelligence (AI) is steeply increasing, specifically in the field of deep learning, resulting in a continuously growing power consumption. Advances in hardware technology will not be sufficient to compensate for this trend. Alternative compute architectures and processing paradigms, such as developed in the field of neuromorphic computing, have the potential for significantly more energy efficient AI solutions. However, such systems are typically limited in size. Even where they have been scaled to large machines, a systematic assessment of power and performance is lacking. This project in the research unit Mod4Comp focuses on the holistic energy and performance modeling of the neuromorphic supercomputer SpiNNcloud, which is a generally-programmable many-core system of 5 million ARM processors connected via a slim, low-latency communication fabric inspired by the human brain. Based on detailed power measurements, energy and performance models will be developed for the SpiNNcloud machine, going from chip level to board and whole-system level. They will be tightly integrated into the Mod4Comp holistic modeling framework. Proxy apps will be developed in close collaboration with the other partners of the research group, reproducing essential properties of application code and forming scalable benchmarks for characterizing the system in terms of power and performance. Models for communication will be refined by dedicated traffic simulation. Spiking neural networks will be the main application focus in this project. Consequently, the developed energy and performance models will be validated in large-scale brain simulations. In the second part of this project, energy and performance models will be used to develop energy-optimized distribution and deployment strategies for large-scale simulation models on the SpiNNcloud. This will have a direct impact on the machine operation, increasing its energy efficiency on system level. The models will further be employed to investigate modified system architectures with e.g. incorporation of near-memory compute approaches. The work in this project will for the first time perform a systematic assessment and modeling of a large-scale brain-inspired compute system. Via integration in the holistic modeling framework of Mod4Comp, a direct and in-detail comparison of the SpiNNcloud architecture to established compute substrates from HPC becomes feasible. This will result in valuable insights for new and more efficient neuro-inspired compute architectures.
DFG Programme
Research Units
