Neural network simulations combine both the data and the hypotheses generated by experimental and theoretical neuroscientists, providing a platform for testing theories and predicting new phenomena. Computational neuroscience continuously challenges existing simulation infrastructure because further biophysical mechanisms get in to focus and there is a demand for ever more complex and diverse models. Large networks with intricate structures and biophysical detail at multiple spatial and temporal scales require accelerated processing and parallel simulation on high-performance computing systems. Simulation codes have lifetimes of decades, while hardware is typically replaced in cycles of years. Leveraging the capabilities of new hardware and software, and at the same time anticipating and addressing the needs of the neuroscientific community necessitates continuous renewal of simulation codes. As the brain itself is a spatially and hierarchically organized system, it becomes another challenge to map the brain architecture onto increasingly heterogeneous and modular computer architectures.

The team keeps a long-term view on upcoming computing systems and provides mid-term coordination for adapting simulation codes to the next computing systems. We identify challenges of emerging technologies and explore new algorithms and proof-of-concept implementations for performance optimization. Systematic performance benchmarks help to ensure that the simulation infrastructure remains efficient and effective, adhering to principles of research software engineering (RSE). Simulations on conventional high-performance computing systems serve as a reference for neuromorphic platforms. As most present supercomputers and upcoming exascale machines rely on GPUs, our current focus is to contribute to the development of the GPU component of the open-source simulator NEST.

Johanna Senk is also a lecturer in Computer Science at the University of Sussex (United Kingdom) and associated with the Sussex AI Centre of Excellence.