̽»¨¾«Ñ¡

Abstract:

Performance benchmarks already drive AI research. Computational neuroscientists investigate brain-like computing by constructing and simulating network models that integrate insights about the structure and function of the brain. The local cortical microcircuit is considered a universal building block of the brain and its characteristics play an essential role in the development of simulation technology for neuroscientific research.
Potjans & Diesmann (2014) proposed a computational model that 1) represents the biological mechanisms of this circuit at realistic spatial and temporal scales, and 2) is sufficiently abstract to allow for analysis via mean-field theory and routine simulations due to only moderate software and hardware requirements. This model has evolved into a standard benchmark for comparing the performance of different simulators in terms of accuracy, time-to-solution, and energy-to-solution.
The simulation results can only be compared on a statistical level due to inherent differences between simulators regarding algorithms, numerical resolutions, or random number generators. In 2018, we compared the simulation performance of NEST running on CPUs with the neuromorphic hardware system SpiNNaker. Although we eventually achieved a good match between the simulation results, at that time, neither technology enabled real-time simulation, and the required power exceeded the demands of the natural brain by orders of magnitude.
Our study was soon picked up by others and subsequent simulations of the same model using and advancing different technologies (including GPUs and FPGAs) has brought a perfomance gain for the community: In only a few years, creative algorithmic strategies have been developed for making best use of the respective systems, and the milestone of real-time has been reached and surpassed at a significantly reduced energy consumption. We consider the benchmarking endeavors around the microcircuit model as a starting point for intensifying the co-development of simulation technologies and increasingly complex yet informative neuroscientific models.

Bio:

Dr. Johanna Senk is a Lecturer in Computer Science at the School of Engineering & Informatics, ̽»¨¾«Ñ¡, UK. She is also the leader of the team “Future Simulation Architectures” at the Institute for Advanced Simulation (IAS-6), Juelich Research Centre, Germany. Dr. Senk studied Physics with a focus on Solid-State Physics and Computational Physics at RWTH Aachen University, Germany, and the University of Trieste, Italy. She received her Dr. rer. nat. degree from the faculty of Mathematics, Computer Science and Natural Sciences of RWTH Aachen University. She performed the research for her PhD and Postdoc in the interdisciplinary field of Computational Neuroscience at Juelich Research Centre. Her research is concerned with uncovering principles of neural network connectivity, including spatial organization and structural plasticity, and novel simulation technology.

By: Aleks Kossowska
Last updated: Thursday, 3 October 2024