Flexible Low Power CNN Accelerator for Edge Computing with Weight Tuning

Abstract

To support various edge applications, a neural network accelerator needs to achieve high flexibility and classification accuracy within a limited power budget. This paper proposes a weight tuning algorithm to improve the energy efficiency by lowering the switching activity. A flexible and runtime-reconfigurable CNN accelerator is co-designed with the algorithm and demonstrated with a feature extraction processor on an FPGA. The system is fully self-contained for small CNNs and speech keyword spotting is shown as an example. A fully integrated custom ASIC is also being fabricated for this system. Based on post place-and-route simulation of the ASIC, the weight tuning algorithm reduces the energy consumption of weight delivery and computation by 1.70x and 1.20x respectively with little loss in accuracy.