Weight Isolation-Based Binarized Neural Networks Accelerator

摘要

In this paper, we introduce a binary neural network accelerator which using a new binarization method and hardware sparse. We propose the weights and the activations to either 1 or 0 instead of +1 or −1, which makes the convolution process simplified and more suitable for hardware implementation. To decrease the data access from off-chip memory, we propose a novel data reuse method, which can reduce 58.8% data access, while the weight isolation logic is designed to reduce power consumption. Based on the weight isolation and the retiming technique, the proposed BNN accelerator achieves low power consumption at 500MHz clock by the VC709 Evaluation Kit. Experimental results show that the proposed accelerator achieves a throughput of 3378 GOPS and 1624 GOPS/W energy efficiency.