The requirement to repeatedly move large feature maps off-and on-chip during inference
with convolutional neural networks (CNNs) imposes high costs in energy and time. In this …