NP-STO / What to look for
Storage is a workload decision.
Throughput numbers below for training come from the NVIDIA SuperPOD with NVIDIA DGX B300 reference architectures. RAG and NIM notes follow how those stacks actually move bytes, plus the NVIDIA NIM support matrix for profile sizing.
Training / post-training
Sustained sequential GB/s and checkpoint burst, not IOPS marketing.
Shared parallel filesystem for datasets. Separate checkpoint target that can absorb a simultaneous flush from the job. NVIDIA SuperPOD B300 RA states high-performance storage I/O per node must exceed 40 GB/s (Ethernet RA) and, on the Quantum-X800 RA, 80 GB/s.
Typical: HPS on IB (QM9700) or RoCE (SN5600) as the storage/data fabric, plus a quieter NFS/user tier on SN5610 as North-South / front-end. Do not fold HPS into a generic East-West compute line if dedicated.
Avoid: A single enterprise NAS asked to be both home directories and the training scratch.
RAG / enterprise retrieval
Random read IOPS, metadata, and a clean split between documents and the vector index.
Object or file store for source documents. NVMe-backed or in-memory vector index. Ingest path sized to nightly rebuilds, not to training checkpoints.
Typical: Object + vector engine. North-South / front-end Ethernet is usually enough if the index is local or on a low-latency tier. That is not the East-West compute fabric.
Avoid: Putting the vector store on the same parallel FS as a training job.
Inference / NIM serving
Fast model-weight load, local NVMe or a warm cache, isolation from training I/O, and a KV / prefix cache plan on large serving.
NVIDIA NIM selects a profile from its support matrix (precision, tensor-parallel size, LoRA). Weights load from local NVMe. That is not the whole inference memory story. On a large workload, KV cache needs a hierarchy: GPU HBM, host overflow, and sometimes a shared prefix cache across the serving plane. That tier is not a storage array and not a checkpoint pool.
Typical: Node-local NVMe for NIM images and weights. KV in HBM first. Host memory and a serving-plane prefix cache when context and concurrency fill the GPU.
Avoid: Loading 70B-class weights over a contended NFS home share. Treating KV as another LUN on the training filesystem.
Research / HPC mix
Scratch that can be purged, project space that cannot, and a scheduler that knows the difference.
Burst scratch on the HPS fabric. Project datasets on a capacity tier. Home and logs on the user-storage Ethernet path, as in the NVIDIA SuperPOD split.
Typical: Two storage systems: HPS + user storage. NVIDIA documents this split explicitly.
Avoid: One quota, one protocol, every lab on the same queue.