A Fully-Configurable Digital Spiking Neuromorphic Hardware Design with Variable Quantization and Mixed Precision
Shadi Matinizadeh, Arghavan Mohammadhassani, Noah Pacik-Nelson, Ioannis Polykretisl, Abhishek Kumar Mishra, James Shackleford, Nagarajan Kandasamy, Eric M. Gallo, Anup Das
Abstract
We introduce QUANTISENC, a fully-configurable digital spiking neuromorphic hardware to optimize performance and power consumption of spiking neural networks (SNNs). QUANTISENC introduces two key contributions. First, it allows the user to set separate quantization and precision policies for the synaptic weights and the internal state variables of neurons to optimize the design based on the precision needed for a target SNN model and the dataset used for training. This reduces the quantization error. Second, in addition to using static design parameters, QUANTISENC also allows to dynamically configure neuron parameters via programming its configuration registers. This allows the user to fine-tune performance and power consumption even after a design is implemented on silicon. Using open-source datasets, we show improvement in area, power, and performance over several state-of-the-art designs.