NVIDIA Opens cuFile API to Speed Up AI Processing with Faster Storage

NVIDIA has announced that it will be open-sourcing its cuFile API and accompanying storage software stack. This move is part of the Storage-Next initiative, which aims to standardize storage systems for faster data transfer between SSDs, network storage devices, and video card memory.

The cuFile technology allows data to be transferred directly between these devices, bypassing intermediate buffered copying through the central processor and system RAM. This reduces latency to microseconds, making it ideal for processing large context windows and AI models.

NVIDIA has partnered with other companies, including Google, Intel, Meta, and Micron, to develop storage system standards. The project will include a universal cuFile API, standards-compliant benchmarks, the multi-platform libxFile library, and the xioLinux database.