To understand why one large metal buffer can be worse than multiple overlapping views, we have to look at how **GPU memory