Highly Efficient 8-bit Low Precision Inference of Convolutional Neural Networks with IntelCaffe