Why can a multi-GPU job wait while GPUs are idle?