What are the primary performance overheads associated with GPU-to-CPU memory transfers, and how do you mitigate them?
