Long-term pulse stimulation and recording in an accelerated neuromorphic system

Accelerated neuromorphic systems allow for fast emulation of spiking neural networks. In these systems, the transmission and handling of pulse events is a demanding task due to the required high bandwidth. In this demonstration, we show the pulse handling in the BrainScaleS waferscale system, which is a large-scale system of this kind. The acceleration of the system requires independent experiment control by dedicated high-speed FPGA network nodes. These nodes release pre-stored stimulus pulses from a playback memory and record neural activity to a trace memory. The external DDR2 memory employed for this task on each FPGA board can hold up to 5 · 10⁸ pulse events, enabling long-term experiments. In this work, we use a fast execution of long emulation times for fine-grained characterization of neuromorphic neurons and small spiking networks. Our results point to the main benefit of accelerated neuromorphic hardware systems, making emulations of long-term experiments feasible, e.g. for learning or behavioural studies.