Nvidia’s response is the shady one, in their response they’re using a different batch size