NVIDIA-Certified Storage#
Introduction#
AI factories present distinct storage demands that traditional benchmarks fail to capture. Standard storage tests often evaluate isolated metrics, such as peak throughput, latency, or input/output operations per second (IOPS), under controlled conditions. However, AI workloads require sustained performance across diverse access patterns. For example, model training demands high-bandwidth sequential streaming and rapid checkpointing, while inference relies on low-latency access to model weights, token embeddings, and key-value (KV) cache states. Production storage must provide sustained and predictable performance, stable latency, reliable data paths, recovery, and consistent operation across the full AI infrastructure as workloads scale.
The NVIDIA-Certified™ Storage program is NVIDIA’s AI Factory data architecture advisory and readiness program for partner storage platforms. NVIDIA works directly with storage partners to assess their architecture, identify operational gaps, and evaluate AI workload performance under stress, scale, and operational demands. The program enables vendors to demonstrate production readiness of their solutions and gives enterprises the confidence to deploy scalable, high-performance storage for AI infrastructure.
How the Program Works#
The NVIDIA-Certified Storage program is organized into three phases.
Architecture Advisory and Readiness#
The engagement begins with architecture advisory and a review of the platform’s readiness for an AI Factory. NVIDIA and the partner review the product architecture, supported AI workloads and frameworks, data services, client and network data paths, scale, availability, redundancy, recovery, security, operations, support model, and product roadmap.
The applicable AI Factory Data design guide establishes the requirements for the engagement. The Storage Platform Assessment and AI Factory Readiness Checklist document how the platform addresses those requirements and provide supporting information for the direct assessment with NVIDIA.
This work identifies product, architecture, configuration, resiliency, support, and operational gaps before production workload qualification begins. Any required capabilities, configurations, patches, or fixes must be incorporated into the partner’s supported generally available (GA) product. The exact configuration used for qualification must be based on that product release.
Production Workload Qualification#
After NVIDIA approves the platform and configuration to proceed, NVIDIA evaluates it at an AI Factory Production Site as a production customer. NVIDIA runs real-world AI workloads to check sustained and predictable performance, scale, stability, p50-to-p99 latency behavior, data-path operation, failure, and recovery.
An important aspect of qualification is the partner’s readiness to support customer operations. By acting as “Customer Zero,” NVIDIA exercises the partner’s ability to respond to problems that can arise during testing, identify and investigate issues, and resolve them in a timely manner. When the evaluation identifies an issue, the partner is expected to engage its product and support teams, incorporate the required correction into its supported generally available product, and work with NVIDIA to revalidate the corrected configuration.
Certification#
Certification is the culmination of this rigorous and collaborative process. NVIDIA grants certification to the exact product and configuration that it tested and approved. Certification is not transferable to a product family, brand, reseller, or another solution.
Foundation, Enterprise, and NVIDIA Cloud Partner (NCP) are progressive certification levels. The following sections describe these levels. Foundation is required before Enterprise, and Enterprise is required before NCP. At each level, the platform must complete the required AI production workloads and meet the acceptance criteria before NVIDIA grants certification.
Benefits of the Program#
For Enterprise Customers: The NVIDIA-Certified Storage program accelerates AI infrastructure deployment. Organizations can use NVIDIA as “Customer Zero” to bypass lengthy acceptance procedures and streamline configuration for storage solutions. It also reduces the financial and time risks of AI initiatives by eliminating hidden integration surprises and exposing operational risks before certification. This gives customers greater confidence in architecture and deployment decisions.
For Storage Partners: The program provides a structured architecture advisory, readiness, and production workload qualification process. NVIDIA works directly with the partner to identify platform gaps, validate supported fixes, and establish a certified configuration and reference design for production AI Factory deployments.
Certification Journey#
Partner engagement begins with AI Factory architecture advisory and readiness. NVIDIA and the partner use the applicable AI Factory Data design guide, Storage Platform Assessment, AI Factory Readiness Checklist, architecture and roadmap reviews, and technical sessions to evaluate the platform and define the exact configuration that can proceed to qualification.
The partner also provides a business case describing the mutual value, market opportunity, target AI Factory use cases, and proposed go-to-market approach. Production workload qualification begins only after NVIDIA and the partner confirm that the product is generally available, the exact configuration is supported, the support model is ready, and identified gaps have been addressed in the partner’s supported GA product.
Foundation Level#
Foundation is the entry certification level. The partner builds, operates, and maintains an NVIDIA-approved Foundation qualification lab. NVIDIA provides a workload qualification package, which the partner runs in the approved configuration and submits to NVIDIA with the complete evidence package.
The qualification package uses AI I/O profiles to evaluate production behavior across file and object data paths. It tests performance across node counts, thread counts, I/O transfer sizes, access modes, and repeated runs. The purpose is to understand scaling, latency, consistency, and data-path behavior across the approved configuration, not to produce a product benchmark or a single peak score.
Each configuration runs multiple iterations to characterize consistency rather than select the best result. AI infrastructure operates continuously, so a platform that reaches the required throughput once but degrades under sustained pressure is not considered ready for production. The evaluation identifies scaling limits, tail-latency growth, namespace contention, and data-path bottlenecks that can appear only as concurrency and system utilization increase.
The qualification package also captures CPU utilization because excessive storage-related CPU consumption reduces the resources available to the AI workload. If training workers wait for checkpoint writes, GPUs can sit idle. If inference replicas wait on small reads or model-shard access, application response time can increase. File and object access paths are evaluated directly as part of the production configuration.
NVIDIA grants Foundation certification after reviewing the submitted results and supporting evidence, confirming that the platform meets the Foundation acceptance criteria, and approving the exact configuration.
Foundation certification validates performance for deployments that use up to 256 GPUs and are based on the following NVIDIA Reference Architectures.
NVIDIA RTX PRO AI Factory.
NVIDIA HGX AI Factory.
Enterprise Lab Readiness#
Enterprise Lab Readiness is a transition between Foundation certification and Enterprise production workload qualification. It is not a separate certification level and does not replace Enterprise qualification at NVIDIA.
After completing Foundation certification, a partner pursuing Enterprise installs the Enterprise-sized storage configuration in the approved Foundation lab. The partner uses the approved Foundation compute clients, networking, and software environment to test the storage hardware and configuration intended for further certification. If testing identifies issues, the partner investigates them, incorporates any required capabilities, configurations, patches, or fixes into its supported generally available product, and revalidates the corrected configuration.
After NVIDIA reviews and approves the results, the partner ships the same approved Enterprise storage configuration and supporting components to the AI Factory Production Site.
Enterprise Certification#
Enterprise builds on completed Foundation certification and NVIDIA approval of Enterprise Lab Readiness. It validates deployments that use up to 4,000 GPUs and are based on the following NVIDIA Reference Architectures.
NVIDIA RTX PRO AI Factory.
NVIDIA HGX AI Factory.
NVIDIA NVL72 AI Factory.
After the approved configuration arrives at the AI Factory Production Site, the partner supports installation and final configuration. NVIDIA operates the environment as a production customer and evaluates the agreed configuration with real AI frameworks, production operating scenarios, and supporting synthetic AI I/O workloads.
The Enterprise evaluation covers performance, scale, stability, p50-to-p99 latency behavior, data-path operation, failure and recovery, configuration quality, and operational behavior. It also evaluates the partner’s ability to investigate findings, deliver supported generally available fixes, and support the qualified configuration as it would any other production deployment.
NVIDIA grants Enterprise certification only after the platform completes the required AI production workloads, sustains the required performance, resolves all blocking issues, revalidates any required supported GA fixes, and meets the Enterprise acceptance criteria.
NVIDIA Cloud Partner Certification#
NVIDIA Cloud Partner (NCP) certification builds on completed Enterprise certification by extending the scale to deployments on the order of 100,000 GPUs. The tests at this level verify the requirements for multitenant AI services, including security, tenant identity and isolation, networking, multitenancy, quality of service, noisy-neighbor control, availability, uptime objectives, operational readiness, and service-level commitments.
Before production workload qualification begins, the partner must expand the certified Enterprise storage platform live to the approved NCP configuration while preserving existing data. NVIDIA grants NCP certification only after the platform completes the required NCP production workloads and operational scenarios, resolves all blocking issues, revalidates any required supported GA fixes, and meets the NCP acceptance criteria.
NCP is intended for storage platforms supporting AI cloud providers and other cloud-scale operating models. Not every Enterprise-certified platform is expected to pursue NCP certification.
AI Production Workload Qualification#
NVIDIA-Certified Storage evaluates storage platforms using AI production workloads and operating conditions appropriate to each certification level. The objective is not to produce a benchmark score or identify a single peak-performance result, but to determine whether the platform can deliver sustained, predictable, and recoverable performance while supporting the data paths, operational requirements, and service expectations of a production AI Factory.
AI production workloads include training, inference, agentic AI, KV cache, checkpointing, model loading, data preparation, and other applicable AI data workflows. NVIDIA can also use supporting synthetic AI I/O workloads to isolate data-path behavior, reproduce issues, and validate corrective actions.
The evaluation considers sustained throughput, scale, stability, p50-to-p99 latency behavior, configuration quality, recovery, and operational response. When an issue is identified, NVIDIA works with the partner to investigate the cause, incorporate required corrections into the supported generally available product and support model, and validate the corrected configuration.
The sections below provide more detail for each workload.
Model Training#
Training large-scale deep learning models requires sustained throughput and parallel I/O across multiple nodes. Storage performance directly affects GPU utilization, training cycle time, and operational efficiency. If the storage layer cannot keep pace with the compute layer, GPUs can idle while waiting for data, reducing overall training efficiency and increasing cost.
Checkpointing is a critical storage operation during training because model state must be saved periodically for recovery and job resumption. Synchronous checkpointing pauses training until checkpoint data is written to storage, requiring high write throughput to minimize downtime. Asynchronous checkpointing reduces disruption by first copying model parameters to CPU memory and then writing the checkpoint to storage in the background. This can lower immediate write-bandwidth pressure and allow checkpoint data to be transferred over a longer period without halting training.
Model Fine-Tuning#
Fine-tuning adapts pretrained foundation models to specific tasks or domains using custom datasets. Although it is generally less resource-intensive than full-scale training, it still requires fast access to model checkpoints and training data, along with predictable writes for updated weights, adapter data, and checkpoints. Storage performance affects iteration time and the ability to support frequent experiments and model updates.
Inference#
AI inference pipelines, particularly for large language models (LLMs) and multimodal architectures, require extremely fast and consistent access to critical data components such as model weights, token embeddings, and feature data. These assets are accessed frequently in real time during inference, making high storage throughput and predictable low-latency performance essential for overall system responsiveness and user experience.
As these inference workloads scale to support real-time applications or multiple users, the demands on the underlying storage infrastructure intensify. The storage system must provide sufficient bandwidth and predictable low-latency performance as compute resources, concurrent requests, model size, and workload complexity increase.
Key-Value Cache#
The key-value (KV) cache is a foundational optimization in transformer-based models, especially for autoregressive tasks like text generation. It stores intermediate key and value tensors from previous forward passes, allowing the model to reuse these computations instead of recalculating them for each new token. This dramatically reduces latency and compute load, which is especially important when serving large models with long context windows or in real-time applications. By using the KV cache, systems can maintain high throughput and responsiveness without sacrificing model performance.
When a workload externalizes or persists KV cache state, storage can support session continuity, cache reuse, and reduced recomputation. The workload test evaluates the applicable KV cache data path and storage behavior for the qualified configuration rather than assuming that every inference deployment persists KV cache to storage.
Query and Retrieval#
Retrieval-augmented generation (RAG) pipelines and agentic workflows enhance large language models by connecting them to enterprise-specific knowledge stored as vector embeddings in high-performance vector databases. During a query, relevant documents, structured data, and contextual information are retrieved through semantic search and provided to the model, enabling more accurate, context-aware responses grounded in an organization’s proprietary data.
To support these workflows at scale, the underlying storage infrastructure must provide low-latency, high-throughput access to vector indexes while maintaining predictable performance under concurrency. As enterprise knowledge evolves, storage platforms must also support secure access controls, encryption, embedding versioning, and efficient index updates without disrupting service. Together, these capabilities help keep semantic search responsive, scalable, and aligned with governance requirements.
Agentic AI Workflows#
Agentic AI workflows introduce a distinct class of storage behavior driven by real agent runtimes. As agents plan, reason, call tools, and produce outputs, they continuously interact with storage to checkpoint session state, restore prior context, append execution traces, update memory, write intermediate tool results, and generate final artifacts. These operations are often small, frequent, and latency-sensitive. When many agent sessions run concurrently, this storage activity can become part of the application’s critical path and directly affect responsiveness, throughput, and reliability.
Reliable storage for agentic AI requires more than high aggregate bandwidth. Agent frameworks often depend on consistent access to state, memory, logs, and artifacts across both isolated and shared namespace patterns. Validating agent framework I/O helps determine whether a storage platform can sustain these framework-visible operations as agent sessions scale, while preserving predictable performance, correctness, and availability across the agentic application layer.
Final Outcome of Certification#
After certification at each level, the partner publishes an NVIDIA-reviewed reference design. The reference design documents the certified configuration, certification level, applicable design guidance, deployment assumptions, sizing considerations, and operating guidance. This gives customers and AI Factory operators a validated and repeatable foundation for production deployment.
The partner can choose to publish one reference design document per certification level or publish a single document with separate sections for each certification level.
Continued Engagement and Recertification#
As is typical with advanced technologies, hardware components, firmware, and software of storage solutions are regularly updated, while NVIDIA continually releases new architectures for GPUs, CPUs, and networking. Additionally, as the AI landscape evolves, new paradigms and algorithms emerge. Ongoing engagement and periodic recertification are therefore essential to ensure storage solutions continue meeting the demanding performance and functional requirements of AI factories over time.
NVIDIA regularly assesses each certified storage solution together with relevant updates to NVIDIA-Certified Storage certification, such as qualification workloads, scripts, and test infrastructure. Based on this evaluation, NVIDIA determines the validation, remediation, documentation update, or recertification activities required for the solution to retain certified status. In all instances, NVIDIA collaborates closely with partners to assess the impact of changes and determine the appropriate recertification path, ensuring sustained reliability for enterprise AI production environments.
Conclusion#
For enterprise technology leaders, certified storage provides more than technical validation. It offers an evidence-based foundation for infrastructure decisions by showing that a platform has been validated for production use in real-world AI deployments. Just as important, ongoing review and recertification help ensure that validated designs remain aligned with evolving operational, performance, and governance requirements. Certification can reduce integration uncertainty, simplify architecture choices, and give buyers greater confidence that the storage layer can support scalable, business-critical AI workloads.