or when you have stronger real-time constraints than "send it to the GPU and hope it comes back fast enough" (e.g. real-time audio processing)