I’m using example transformer models from huggingface which can be found in here: https://github.com/huggingface/transformers/tree/master/examples/tensorflow/question-answering. Simply follow the README.md, you can see that M1 max is about 1/6 the training speed of RTX 3090.