Abstract
With more and more event-based neuromorphic hardware systems being developedat universities and in industry, there is a growing need for assessing theirperformance with domain specific measures. In this work, we use the methodologyof converting pre-trained non-spiking to spiking neural networks to evaluatethe performance loss and measure the energy-per-inference for threeneuromorphic hardware systems (BrainScaleS, Spikey, SpiNNaker) and commonsimulation frameworks for CPU (NEST) and CPU/GPU (GeNN). For analog hardware wefurther apply a re-training technique known as hardware-in-the-loop training tocope with device mismatch. This analysis is performed for five differentnetworks, including three networks that have been found by an automatedoptimization with a neural architecture search framework. We demonstrate thatthe conversion loss is usually below one percent for digital implementations,and moderately higher for analog systems with the benefit of much lowerenergy-per-inference costs.