Controlling which data stays in GPU memory based on when it's produced and consumed across pipeline stages.