Future large language models will require orders of magnitude more memory bandwidth to run faster