[cubefox]

That's the question. More precisely, how does the new method compare to the classical one in terms of training compute and inference compute?