You compare inference computation, but the majority of resources are applied at training time. I believe the real number for hardware might be closer to the result you got for bees.