NetApp is targeting one of the less visible constraints in artificial intelligence infrastructure with Novus, a new storage architecture designed to keep data moving fast enough to support large-scale GPU clusters.
The company said GPU utilisation can fall below 30% when storage infrastructure cannot deliver data at the speed required by AI workloads, potentially leaving expensive computing capacity underused.
Novus addresses this by separating file metadata from data movement, allowing storage capacity and bandwidth to be scaled independently rather than tying the two functions together.
Its initial configuration combines Novus Data Director running on Supermicro infrastructure with ONTAP services on NetApp AFF A90 systems.
NetApp said the architecture is designed to support AI environments spanning hundreds of thousands of GPUs and is now available to order.
The company is positioning the technology for so-called AI factories, where large GPU clusters continuously process data for training and running AI models.
“AI factories struggle and GPU economics collapse when data can’t keep up,” said NetApp chief product officer Syam Nair.
Omdia’s testing of Novus showed near-linear performance gains as additional ONTAP clusters were added. Its modelling projects sequential read throughput of up to 100 terabytes per second, although NetApp did not report this as the result of a full-scale deployment.
The launch comes as AI infrastructure shifts towards increasingly large clusters, putting greater pressure on the storage systems feeding data into GPUs. While much of the focus in AI infrastructure has centred on chips and computing capacity, storage performance can also become a limiting factor as workloads scale.
NetApp also introduced AI-assisted operations in its NetApp Console, allowing organisations to manage data across on-premises infrastructure and public cloud environments.
The company separately launched Keystone Sovereign for qualified customers in the European Economic Area and said an integration with Nutanix Cloud Infrastructure for virtualised workloads is scheduled to become generally available later this year.
