Scaling up BERT-like model Inference on modern CPU - Part 2