For example, he said that with the default PyTorch technology, training an 11-billion-parameter model, over an ethernet-based network, could be done with only 20% GPU efficiency.