ABSTRACT
This short article explains why the Epiphany architecture is a proper reference for digital large-scale neuromorphic design. We compare the Epiphany architecture with several modern digital neuromorphic processors. We show the result of mapping the binary LeNet-5 neural network into few modern neuromorphic architectures and demonstrate the efficient use of memory in Epiphany. Finally, we show the results of our benchmarking experiments with Epiphany and propose a few suggestions to improve the architecture for neuromorphic applications. Epiphany can update a neuron on average in 120ns which is enough for many real-time neuromorphic applications.
