Inefficient data loading from storage (B) is the most likely cause of slow performance when storage I/O is consistently high. In AI training, GPUs require a steady stream of data to remain utilized. If storage I/O becomes a bottleneck-due to slow disk reads, poor data pipeline design, or insufficient prefetching-GPUs idle while waiting for data, slowing the training process. This is common in shared clusters where multiple jobs compete for I/O bandwidth. NVIDIA's Data Loading Library (DALI) is recommended to optimize this process by offloading data preparation to GPUs.
* Incorrect CUDA version(A) might cause compatibility issues but wouldn't directly tie to high storage I
/O.
* Overcommitted CPU resources(C) could slow preprocessing, but high storage I/O points to disk bottlenecks, not CPU.
* Insufficient GPU memory(D) would cause crashes or out-of-memory errors, not I/O-related slowdowns.
NVIDIA emphasizes efficient data pipelines for GPU utilization (B).