Automated Code Generation for Deploying NESTML Neuron Models on SpiNNaker2

Zenodo (CERN European Organization for Nuclear Research)

Abstract

Poster presented at ICNCE 2026 Abstract Neuromorphic computing platforms promise scalable and energy-efficient simulation of large spikingneural networks, but deploying neuron models on specialized hardware often requires manual low-level implementation. This creates a barrier between high-level neuron model description languagesand hardware-specific deployment. To surmount this barrier, the domain-specific modeling languageNESTML [1] enables concise specification of neuron models in a user-friendly syntax, combined withautomated code generation for specific simulation platforms, thereby maintaining high performance.In this work, we extend the NESTML code generation pipeline to support the neuromorphic SpiN-Naker2 [2] platform, enabling automatic translation of neuron models into model-specific C-codeimplementations and Python-code interface classes. The approach introduces an intermediate trans-formation stage that eliminates physical unit annotations and converts expressions into dimensionlessnumerical form suitable for embedded execution. In addition to analytic integration, two numericalintegration schemes are implemented to evaluate solver trade-offs on the platform: We provide a For-ward Euler method solver with low computational cost and a fourth-order Runge-Kutta (RK4) solverproviding improved numerical accuracy. Generated implementations are validated against referencesimulations in NEST Simulator [3] using leaky integrate-and-fire and Izhikevich neuron models.The results show that generated implementations preserve expected firing statistics and qualitativenetwork dynamics when appropriate numerical parameters are chosen, based on membrane poten-tial trajectories, spike timing accuracy, and numerical stability across different integration step sizes.Performance measurements indicate that automatically generated code is fractionally slower thanhand-optimized implementations, but significantly reduces development effort and enables rapid de-ployment on SpiNNaker2. In conclusion, SpiNNaker2 support for NESTML is a key enabler in thetransition towards power-efficient neuromorphic computing.[1] C. Linssen et al., Front. Neuroinform. Volume 19:1544143, 2025[2] H. Gonzalez et al., Machine Learning with New Compute Paradigms Conference, 2023[3] M. Diesmann and M.-O. Gewaltig, Forschung und wisschenschaftliches Rechnen, Beiträge zum Heinz-Billing-Preis,Volume 58, Pages 43-70, 2001

Authors 5

1. Martin Weih

   Technische Universität Dresden

   Affiliation as printed

   Technische Universität Dresden

2. Charl Linssen Aachen

   RWTH Aachen University · Forschungszentrum Jülich

   Affiliation as printed

   Forschungszentrum Jülich

   RWTH Aachen University

3. Bernhard Vogginger

   Technische Universität Dresden

   Affiliation as printed

   Technische Universität Dresden

4. Abigail Morrison Aachen

   RWTH Aachen University · Forschungszentrum Jülich

   Affiliation as printed

   Forschungszentrum Jülich

   RWTH Aachen University

5. Christian Mayr

   Technische Universität Dresden

   Affiliation as printed

   Technische Universität Dresden