#data-loading

Question 1: Single choice

Direct memory access (DMA) lets a device such as a GPU, a network card or an SSD read and write main memory without the CPU copying each word. Why is this useful when training a model?

The CPU is free for other work, such as preparing the next batch, while the transfer runs1
The data travels over a private link instead of PCIe, so transfers beat the bus bandwidth2
Transfers become synchronous, so the CPU knows exactly when each copy has finished3
The device compresses the data while copying it, so datasets take up less memory4

Timed test