Question

A data science team is training a computer-vision model on a DGX system. During training they observe that GPU utilization stays around 40%, while their monitoring shows the small local disk is saturated reading millions of image files sequentially through the CPU. They want to redesign the storage layer so that many nodes can read the same large dataset concurrently at high aggregate throughput. Which storage approach best addresses this requirement?

Accepted Answer

A high-throughput parallel file system (e.g., shared NVMe-backed scale-out storage) accessed over a high-speed network by all nodes

Answer

A single large SATA hard disk drive attached directly to one node and shared via a slow NFS export

Answer

Copying the full dataset into each GPU's VRAM before every training epoch

Answer

Storing the dataset in the CPU's L3 cache to reduce read latency