Performance and energy modeling for spiking network simulations on conventional and accelerator architectures

• Applicant: Dr. Susanne Kunkel
• Subject Area: Computer Architecture, Embedded and Massively Parallel Systems ; Software Engineering and Programming Languages
• Term: since 2026
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 545776403

This project in the research unit Mod4Comp advances performance models of simulation technology for spiking neuronal networks on conventional computer architectures with a focus on many-core systems and energy efficiency. In recent years the community has achieved a separation between generic simulation engines and concrete neuronal network models: many models, in particular also future ones, can be simulated with the same simulation engine. This enables the operation and maintenance of such a generic code as a scientific infrastructure. The NEST code is the leading simulation tool for spiking neuronal networks at the resolution of neurons and synapses. It is designed for the study of neuronal networks at their natural size up to billions of neurons, using the largest available supercomputers including the upcoming exascale systems. In addition NEST is established as the reference for neuromorphic computing systems and models first implemented in NEST have become de facto standards for benchmarking. The dynamics of spiking neuronal networks is characterized by highly irregular sparse activity on a sparse graph. Combined with the memory consumption caused by the several thousands of incoming synapses of a neuron this challenges the memory bandwidth and cache efficiency of processors. The NEST code has iteratively been optimized over the years and preliminary profiling data show that there still is substantial potential for further reduction of memory latency and fine-grained parallelization. Recent optimization efforts have also targeted GPUs and explored the potential of massively distributed memory on Graphcore's IPU. Overall, further progress, however, relies on the availability of detailed performance models. These models will guide optimizations for existing hardware with respect to time-to-solution and energy-to-solution in the subsequent second phase of the project and enable predictions for future systems including neuromorphic accelerators. An initial objective is to dissect the simulation cycle into phases and to capture the flow of spikes in state-of-the-art code by a model. In parallel a small suite of real-world neuronal network models covering the application domain of the simulation code will be constructed. The project will contribute to a Library of Proxy Apps highlighting critical sections of the code together with other projects while the performance and energy modeling will be done in close collaboration with other projects. The results provide feedback to all levels of the multilayer modeling module of the research unit. Data on memory performance enter the considerations on memory architectures of SP1. Finally, SP4 will work closely with SP6 to set simulation speed and energy consumption of a neuronal network using the full SpiNNaker2 installation into perspective with past achievements and present supercomputers.

• DFG Programme: Research Units
• Subproject of: FOR 5880: Holistic Energy and Performance Modeling for Sustainable Computing (Mod4Comp)
• Co-Investigator: Professor Dr. Markus Diesmann