Is it just memory bandwidth? Or is it that AMD is not well supported by pytorch well enough for most products? Or some combination of those?