Decoding speed won't matter one bit if you have to sit there for 5 minutes waiting for the model to ingest a prompt that's two sentences long.