In NV GPU, UM mode means when you put your data in "managed memory", then GPU and CPU can use the same virtual address pointer to access that data in host memory. This is actually done by page copying between CPU and GPU, but the effect is that CPU and GPU are accessing the same piece of data with same VA pointer. This is convenient but the problem is that the performance is much lower in UM mode because it takes a lot time to do the page copying and managing (disable or enable page tables on CPU and GPU).
So my question is: who is really using UM mode with GPU in real application? what's the typical usage scenarios?