KUALA LUMPUR: NetApp has unveiled a storage architecture designed to keep data flowing to large clusters of graphics processing units (GPUs), addressing a bottleneck that can leave costly AI infrastructure underused.
Called NetApp Novus, the architecture separates file metadata management from the movement of data. NetApp said this would allow operators to expand storage capacity and bandwidth independently while keeping files accessible through a single namespace.
The company said GPU utilisation in AI factories can fall below 30% when storage systems cannot supply data fast enough. The problem is particularly costly for providers that rent out GPU capacity and depend on keeping their equipment in use.
“AI factories struggle and GPU economics collapse when data can’t keep up,” said NetApp chief product officer Syam Nair.
The initial Novus release combines NetApp Novus Data Director software running on qualified Supermicro infrastructure with ONTAP data services delivered through NetApp AFF A90 systems. NetApp said the architecture is available to order and is designed to support high utilisation across hundreds of thousands of GPUs.
NetApp is targeting aggregate throughput of more than 100 terabytes per second as the architecture scales. Omdia chief analyst Tony Palmer said testing observed by the research firm showed near-linear performance gains as ONTAP clusters were added. Omdia’s modelling projects 100 terabytes per second in sequential read throughput and dozens of exabytes of effective capacity.
NetApp said the architecture also provides a path towards zettabyte-scale storage capable of serving millions of GPUs. Those larger-scale figures remain projections, rather than results from a deployed system of that size.