Would You describe how did You split the tasks for particular GPUs such way?