NetApp has introduced a new data architecture designed to maximize efficiency and utilization of millions of GPUs in AI infrastructure environments, the company announced Tuesday during its NetApp Insight conference in Las Vegas. CEO George Kurian called NetApp Novus the company's most significant product launch in a long time, describing it as an intelligently disaggregated system where every component can be optimized for performance and scaled on its own. The architecture aims to support AI factories operating clusters of 100,000 or more GPUs while enabling access to over 1 zettabyte of data through a single namespace.
The new platform delivers performance of 100 terabytes per second, according to Kurian, who said this speed allows researchers to run complex simulations and models in real time or enables biologists to use tens of thousands of computers to monitor every protein molecule in a cell. A single GPU sitting on high-bandwidth memory can process up to 8 terabytes per second of data, but traditional storage infrastructure can't keep pace, said Arindam Banerjee, NetApp's chief platform and technology officer. He told CRN that GPU utilization could drop to as low as 5 percent when the rest of the data infrastructure fails to match that speed. Customers who discussed 10,000 GPUs as state-of-the-art just last year are now planning clusters of 50,000 to 100,000 GPUs, Banerjee said, noting that serving 50,000 GPUs from a namespace at a constant line rate of 2 gigabytes per second requires 100-terabyte-per-second bandwidth.
Kurian said the architecture is built for the accelerated computing era, which brings fresh demands and opportunities for infrastructure and system design. According to Banerjee, the most expensive asset in a data center isn't being utilized when GPU performance falls due to infrastructure bottlenecks, and this challenge keeps evolving as new GPU families arrive each year. The company disaggregates metadata from data because scaling both together on the same infrastructure is extremely difficult, Banerjee explained, noting that metadata involves small transaction I/Os while data involves large sequential I/Os. NetApp federates the namespace across many data clusters while still using open standards like parallel NFS and storing data in OnTap, preserving features like resiliency, end-to-end data integrity, and security that customers expect.
Channel partners will play an important role as the technology runs on NetApp's A90 flash storage systems, which partners already sell to customers, said Jen Prenner, senior vice president of product marketing. Ned Engelke, CTO at Evotek, a San Diego-based solution provider and NetApp channel partner, said his company has several customers with scale they can't address today without a fresh approach, including pharmaceutical companies with 20 or 30 years of sequencing scans stored in hopes of rediscovering them or companies with intellectual property around chip development that might need to uncover data to respond to lawsuits. Engelke noted that customers often have exabytes of data they can't discover or use, forcing them to pay storage costs they regret daily. The technology should help partners produce different architectures for customers who need to distribute data, process it, and keep it secure, he said. For businesses competing on differentiation through data insights, the ability to finally access massive legacy data estates could unlock lessons from past projects and better protection against legal challenges, fundamentally changing what's possible when computational power no longer sits idle.