What about int8 performance? Would it be somehow improved by using tensor cores?