3 ms·

I think this is valid for training but Nvidia also dominates inference, which is a bit of a different argument