Operating Accelerated Neuromorphic Hardware - A Scalable and Sustainable Approach

Accelerated mixed-signal neuromorphic hardware presents a promising approach to overcome run time and scalability issues of software-based neural network simulations. It accomplishes this by physical emulation of the neuronal dynamics via specialized analog circuitry instead of numerical calculations. However, facilitating the advantage of such highly custom hardware with a similar convenience as conventional simulators poses various challenges. This thesis addresses these in two ways: First a multi-layered software architecture developed for the second-generation BrainScaleS neuromorphic systems is presented. Welldefined interfaces allow utilization of the hardware in different stages of development with the appropriate level of abstraction. The upper layers provide an interface to effciently describe neuroscience experiments and handle automated translation of population-based spiking neural network graphs to valid hardware configurations and experiment flow programs. Suitable run time performance and scalability of the software are verified by extensive measurements while usability is demonstrated via an SNN-based Sudoku solver. The second part covers the challenges of supplying novel compute hardware as a research platform to the neuroscience community. A convenient and robust multi-user access is facilitated via customization of the prevalent SLURM resource scheduler to the requirements of neuromorphic experiment workflows. Finally, a monitoring infrastructure vital for system commissioning and experiment reproducibility is established