Itdaily - NetApp Novus keeps hundreds of thousands of GPUs well-fed

NetApp Novus keeps hundreds of thousands of GPUs well-fed

Low-angle view of a glass office building with the NetApp logo on its top facade.

NetApp introduces Novus, a storage architecture for providers with large GPU environments. The system separates metadata from the data path and is designed for a throughput of 100 TB/s.

NetApp has announced Novus, a new architecture for neocloud providers and GPU-as-a-Service companies building AI factories. According to the company, traditional storage architectures fall short at this scale. As a result, GPU utilization can drop below thirty percent because data does not reach the chips fast enough. This reduces the return on expensive investments and increases infrastructure costs.

Metadata separate from the data path

Novus separates metadata from the data path. Consequently, performance, capacity, and concurrent access can scale independently under a single namespace using standard NFS. NetApp is building the system for a file system at zettabyte scale.

The first release combines Novus Data Director, the metadata software, with qualified Supermicro infrastructure. This is complemented by NetApp’s ONTAP data services on AFF A90 systems. The result is a disaggregated architecture that, according to the company, achieves a throughput of more than 100 TB/s and can efficiently serve hundreds of thousands of GPUs.

The architecture is intended for multi-tenant AI factories at cloud scale and is available for order today. In time, a path to software-defined deployment will be introduced, featuring the same namespace and operational model.

At the scale of AI factories

NetApp lists three advantages of this new architecture. By decoupling the metadata, there should be fewer storage bottlenecks, keeping the GPUs working. Billions of files and large AI datasets grow within a single namespace, without silos or additional operational complexity.

Furthermore, operators only scale the resources that a workload requires. This is intended to limit the consumption of power, space, and infrastructure, while maintaining reliability, security, and multi-tenant capabilities.

Syam Nair, Chief Product Officer at NetApp, states that storage from the HPC era alone cannot handle the scale of AI factories. According to him, an architecture is needed that is built for that scale from the ground up.

Scalability tested

Various analysts, including Omdia, have tested the Novus architecture. In these tests, performance scaled almost linearly as ONTAP clusters were added to a single global namespace. Based on this, Omdia expects Novus to scale up to 100 TBps of sequential read throughput, with tens of exabytes of effective capacity. The firm anticipates that the solution can also handle zettascale implementations.

Mike Leone, Vice President and Principal Analyst at Moor Insights & Strategy, points out that the economics of AI infrastructure depend on the utilization rate of computing power. According to him, storage that fails to deliver data at cloud scale limits training performance and return on investment.

Cisco, which provides FlexPod storage together with NetApp, also supports the approach. Jeremy Foster, Senior Vice President and General Manager of Cisco Compute, sees the data and its utilization as the true differentiator, not the GPU itself.

With Novus, NetApp is targeting providers who want to better utilize their GPU investments. The architecture is now available for order, while software-defined deployment will follow later. The figures regarding throughput and scale come largely from NetApp itself and from Omdia’s projections, so practice has yet to prove whether Novus delivers on these promises.