NetApp ONTAP AI with NVIDIA DGX SuperPOD Validated Design

The NetApp ONTAP AI with NVIDIA DGX SuperPOD Validated Design is a jointly engineered, end-to-end reference architecture that integrates NetApp AFF all-flash storage systems running ONTAP data management software with NVIDIA DGX SuperPOD GPU compute infrastructure. The validated design defines a prescriptive, tested configuration in which high-throughput NFS over RDMA (RoCE) or InfiniBand fabric interconnects the DGX compute nodes directly to NetApp AFF arrays, eliminating storage bottlenecks that typically constrain large-scale deep learning training pipelines. NetApp ONTAP provides enterprise-grade data services—inline data reduction, snapshot-based data protection, multi-tenancy, and QoS—on top of the raw performance required to saturate hundreds of GPUs simultaneously.

The architecture is structured around a scale-out building-block model: a base SuperPOD pod unit consisting of NVIDIA DGX H100 or A100 nodes is paired with a defined quantity of NetApp AFF storage nodes and Mellanox/NVIDIA InfiniBand or Ethernet storage fabric switches, with the entire stack validated by joint NetApp and NVIDIA engineering teams. The ONTAP AI validated design documentation specifies exact cabling topology, network configuration, ONTAP storage efficiency settings, and MLPerf-aligned benchmarking procedures, giving infrastructure teams a reproducible, supportable deployment path rather than a best-effort assembly of components.

Target workloads include large language model (LLM) pre-training and fine-tuning, computer vision model development, reinforcement learning from human feedback (RLHF) pipelines, genomics and drug discovery inference workloads, and financial risk modeling. The design supports both on-premises deployment in enterprise datacenters and colocation facilities, and integrates with NetApp's hybrid cloud data fabric, enabling datasets to be tiered or replicated to major hyperscaler object stores while keeping hot training data on low-latency all-flash NVMe media co-located with the GPU cluster.

Ideal for

• Large-scale foundation model and LLM pre-training requiring sustained high-throughput data ingestion across hundreds of NVIDIA H100 or A100 GPUs without storage-induced idle cycles
• Enterprise MLOps pipelines where multiple data science teams require isolated, QoS-governed access to shared GPU-attached storage with snapshot-based dataset versioning and rollback
• Genomics, drug discovery, and life sciences AI workloads demanding simultaneous high-throughput sequential reads of large binary datasets (FASTQ, BAM, DICOM) at training time
• Financial services deep learning applications—fraud detection, algorithmic risk modeling—that require immutable audit-trail snapshots and data-at-rest encryption with FIPS 140-2 validated key management
• Government and defense AI workloads operating in air-gapped or restricted-network datacenters that require a fully on-premises, validated, and reproducible GPU-plus-storage stack
• Media and entertainment generative AI rendering pipelines that produce and consume very large unstructured file datasets and benefit from ONTAP's inline deduplication and compression to reduce usable storage footprint

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.