I very consistently see people say prompt processing is slow for larger context sizes ("notoriously slow"), something that is much less of an issue with eg CUDA setups.