A Circuit Concept for Energy-Efficient Spiking Neural Network Systems with a FOM of 86.9fJ/SOP

This paper shows a circuit concept for low-energy neuromorphic hardware for Spiking Neural Networks (SNNs). Combining continuous time-encoded values with time-discrete circuitry helps to improve energy-efficiency. Simple design blocks for neurons and synapses make a scalable neural network design process possible. The direct connection between neurons and synapses in an analog manner reduces latency and energy consumption. The functionality was proven by a small neural net solving a classification problem on silicon. Measurements show a FOM of 86.9 fJ for one synaptic operation.