Trying to get started and run the pretrained NER pipeline model (https://huggingface.co/transformers/task_summary.html#named-entity-recognition) on about 10 million instances of text. Would you recommend using a CPU or GPU?
Previously, I used spaCy and broke my data up into batch sizes of 1k then ran those batches on a C5 instance on AWS. The AWS EC2 instance was compute optimized and had 96 cores. I’m able to run any instance on AWS. Thanks!