Nvidia open sources cuFile API, accelerating GPU read/write capability for high-speed storage - SiliconANGLE

NVIDIA has open-sourced the cuFile API, a key component of its GPU storage stack, to accelerate data access speeds for high-performance computing. As AI applications demand faster data access, this mo... Read original →

NVIDIA’s open-sourcing of the cuFile API marks a pivotal shift in GPU storage architecture, directly addressing traditional data access bottlenecks. By enabling GPU Direct Storage to move data from NVMe drives directly into GPU memory, it significantly boosts AI inference and training speeds. This innovation pressures downstream vendors like DataDirect Networks, Kioxia, and Micron to accelerate compatibility upgrades and advance 3nm EUV-based storage chips. However, direct GPU-to-storage access introduces security concerns, prompting NVIDIA to adopt SCADA for access control, though this may spark debates over data sovereignty. Competitors like AMD and Intel may respond with proprietary storage interfaces to counter NVIDIA’s dominance. In the longer term, this move will redefine AI infrastructure standards, accelerating the adoption of large-scale models such as mixture-of-experts, and establishing new competitive moats in the semiconductor and AI stack.