問題文
In a training node, what is the primary role of the host CPUs?
選択肢
- To perform the matrix operations of the model, with the accelerators used only for storing the weights between steps, so the core count of the host sets the pace of a step.
- To terminate the network connections of the fabric, because accelerator traffic must always be copied into host memory by the CPU before it can leave the node.
- To read, decode and transform the data set fast enough that the accelerators are never left waiting for their next batch.
- To hold the entire data set in main memory for the duration of the run, on the grounds that a training job cannot read from storage once the first step has started and therefore depends on the host memory as its only source.