Nvidia has open-sourced its cuFile storage stack with Google, Intel and Meta as founding maintainers, aiming to establish an open standard for GPU-native storage that improves performance, Linux interoperability and ecosystem adoption.
Nvidia has open-sourced the cuFile APIs and the complete cuFile storage software stack, marking a major step towards making GPU-native storage access an open industry standard. The code has been released under the new Accelerated IO Special Interest Group (SIG) on GitHub, with Google, Intel and Meta joining Nvidia as founding maintainers despite competing in AI hardware.
The move shifts cuFile from a proprietary Nvidia technology to a community-driven open-source project that storage vendors and AI infrastructure providers can adopt across the industry.
At the centre of the release is SCADA (Scaled Accelerated Data Access), a framework that moves storage control directly onto the GPU. Instead of relying on CPUs to issue storage requests, GPUs can generate requests and process completions themselves, keeping hundreds of thousands of operations in flight simultaneously while reducing storage latency to microseconds. The technology builds on GPUDirect Storage, introduced in 2021, which removed CPUs from the data path but still depended on them for storage control.
According to Nvidia, open-sourcing cuFile also strengthens Linux interoperability and security by using standard Linux privilege mechanisms for direct storage access.
Supporting the announcement, Micron demonstrated the technology on a single server equipped with 44 Gen6 SSDs, three PCIe switches and three Nvidia H100 GPUs, achieving 230 million random 512-byte reads per second—around 95% of the drives’ rated performance—while the CPU remained essentially idle.
Nvidia also launched Storage-Next, an ecosystem initiative involving more than 40 companies, including DDN, Kioxia and Micron, to advance open GPU-native storage technologies.
Jason Hardy, Vice President of Storage Technology at Nvidia, said, “The result is data pulled from storage in microseconds.”
