hard20 pts · first answer countsgpu-computingdata-loadingperformance

Overlapping host-to-device copies

With pinned memory, what does x.to("cuda", non_blocking=True) allow that a normal copy does not?

A.The copy skips the PCIe bus entirelyB.The copy can overlap with GPU compute and CPU workC.The tensor is compressed before sendingD.The copy happens after training ends

Discussion 0

Join free to post in the discussion.

Overlapping host-to-device copies

Discussion 0

Related questions