Nvidia Corporation Open-Sources cuFile APIs to Accelerate AI Storage

Nvidia Corporation open-sourced its cuFile application programming interfaces and vertical storage software stack to reduce AI storage latency. The software enables GPUs to read and write to storage directly, bypassing CPUs to achieve microsecond-level latency. Google LLC, Intel Corporation, Meta Platforms Inc., and Nvidia Corporation serve as the inaugural maintainers of the project.

Concurrent with the software release, the company launched Storage-Next, an initiative comprising more than 40 storage and flash suppliers, such as Micron, Kioxia, and DataDirect Networks. This effort aims to create open industry standards for GPU-driven storage through the SCADA framework, which allows parallel GPUs to pull specific data directly into memory while adhering to Linux-based security protocols.

DataDirect Networks has already begun integrating the SCADA framework into its Infinia software-defined data platform. The move is intended to increase the efficiency of accelerated computing resources by optimizing the connection between GPUs and data.