Speeding up model training involves more than kernel tuning. Data loading frequently slows down training, because datasets are too large to fit on disk, consist of millions of small files, or stream ...