Nvidia wants to bypass the CPU with an open-source AI storage overhaul

Nvidia has open-sourced its cuFile API, enabling GPUs to read and write directly to storage without CPU involvement, with Google, Intel, and Meta joining as core maintainers. The company also released the SCADA vertical storage software stack, and DDN is an early adopter using it to connect Nvidia GPUs to its Infinia platform. DDN CTO Sven Oehme said the collaboration aims to reduce friction between compute and data to improve AI economics.

Nvidia is looking to help speed up AI workloads by open-sourcing an API that lets a graphic processing unit GPU read and write directly to storage without having to source data from the central processing unit CPU . The cuFile https://github.com/xio-sig API was published to GitHub, with Google, Intel, and Meta joining the chip giant as core maintainers. Its open-sourcing comes amid a renewed focus on the CPU given its importance in agentic AI stacks as the central control plane and orchestration engine. CuFile, however, enables the movement of data through a drive to GPU memory without the need for a bounce buffer from the host CPU’s random access memory RAM CuFile essentially allows hundreds-of-thousands of GPU threads to securely access data from storage “in just microseconds.” Nvidia also open-sourced the scaled accelerated data access SCADA vertical storage software stack that sits beneath the API. The chip giant said the move would “help make security context, data and storage accessible at the speed AI-powered defenses need.” The SCADA framework allows parallel GPUs to extract specific data required for applications directly into high-speed memory, optimizing performance by bypassing unnecessary information. DDN https://www.ddn.com/press-releases/ddn-and-nvidia-collaborate-to-advance-gpu-initiated-data-access-for-next-generation-ai/ is an early adopter of the platform and is using it to provide more efficient connections between Nvidia GPUs and the vendor’s Infinia data intelligence platform. “AI success will increasingly be measured not by how much infrastructure an organization owns, but by how productively it uses that infrastructure," DDN CTO Sven Oehme explained. “Our work with Nvidia is focused on helping customers keep their GPUs working, accelerate time to insight, and improve the economics of AI at scale. By reducing the friction between compute and data, we can help organizations generate greater value from every accelerator, watt, and dollar invested.”