4 ms·

For inference it makes a lot of sense, but for training its going to be significantly slower then buying 2 3090s