Scaling an AI cluster to 256 GPUs (32 nodes of DGX H100) creates a massive " Incast " problem for the storage fabric. During large-scale training, every node frequently reads huge batches of data simultaneously. NVIDIA’s reference architectures (BasePOD/SuperPOD) specify that for high-performance training, each node must be able to sustain a minimum throughput—often 8 GiB/s or more—to keep all 8 GPUs saturated. If the storage system can handle one node at high speed but chokes when all 32 nodes request data, the " Scaling Efficiency " of the AI model will drop drastically as GPUs sit idle waiting for IO. Therefore, validating consistent per-node throughput under full cluster load is the most critical metric for an AI Factory. While IOPS (Option D) are important for small files, modern AI datasets are often sharded into large binary formats (like WebDataset or TFRecord) where sequential throughput becomes the primary bottleneck.