NetApp launches Novus AI storage architecture for large GPU clusters
NetApp says Novus separates metadata from data access to target AI-factory storage bottlenecks, but customer results remain unproven.
By Wei-Lin Zhao · AI Correspondent
· 3 min read
NetApp on Sept. 29 announced NetApp Novus AI storage, an ONTAP-based architecture aimed at AI factories, neoclouds and GPU-as-a-service operators. The system separates metadata handling from data movement, a design NetApp says is intended to prevent storage work from leaving large GPU fleets waiting for data.
The launch is a bid for AI infrastructure projects where data systems must serve both many small file operations and sustained, high-bandwidth training traffic. Cloud infrastructure combines configurable compute, storage and networking resources. Novus addresses the storage portion of that stack; NetApp did not disclose pricing in the supplied materials.
How does NetApp Novus address AI storage bottlenecks?
Metadata is the information a file system uses to locate and describe files. In a large training environment, operations such as opening a file, looking it up or requesting its layout can arrive in large volumes. Those requests can contend with bulk operations such as checkpoint writes when the same controllers handle both jobs, according to NetApp’s technical description.
Novus puts those tasks on separate layers. Its Novus Data Director operates as a software-defined metadata and control layer. A client requests a file location and layout from Data Director, then reads or writes directly to the storage layer. NetApp says that structure lets the metadata layer and the data layer grow independently.
For its initial configuration, NetApp says Data Director runs on qualified Supermicro infrastructure, while ONTAP data services run through AFF A90 systems. The company says the architecture federates ONTAP-based storage under one namespace, so applications see a unified file view as storage systems are added rather than separate mounts. It uses standards-based NFS and pNFS access; NetApp says supported Linux environments can use in-kernel NFSv4.2 and pNFS Flex Files clients.
What scale and availability has NetApp claimed?
NetApp says Novus is designed to exceed 100 TB/s of aggregate throughput and support hundreds of thousands of GPUs. Its illustration assumes roughly 2 GB/s per GPU, which would require 100 TB/s across 50,000 GPUs. Those are design targets and capacity arithmetic, not reported results from a customer deployment.
The company says Novus is orderable now and describes a future path toward software-defined deployments. Its announcement also says statements about unreleased offerings and future plans can change and are not commitments on availability, functionality, pricing or timing.
Established by the announcement: the initial Data Director, Supermicro and AFF A90 configuration, plus NetApp’s statement that Novus is orderable.
Vendor design claims: a single namespace, independently scalable metadata and data layers, and throughput above 100 TB/s.
Not established in the supplied record: realized customer throughput, GPU-utilization gains, cost savings or returns on investment. The supplied announcement and supporting materials also do not establish pricing, customer deployments, availability regions or independent latency benchmarks.
The distinction matters because NetApp is entering a field with established AI-factory storage specialists. The Futurum Group characterized NetApp as a later entrant to that segment, while noting that Novus targets AI-factory builders and NetApp’s AI Data Engine is aimed at enterprise agentic-AI users. NetApp has also announced an intent to acquire PEAK:AIO; the deal terms were not disclosed.
This story draws on original reporting from SiliconANGLE.
