Nvidia Unveils Open-Source Storage Overhaul for AI Workloads

Nvidia has open-sourced an API called cuFile to enable GPU-accelerated storage, allowing GPUs to read and write directly to storage without needing to access data from the central processing unit (CPU). The move is part of a renewed focus on optimizing AI workloads.

CuFile enables hundreds-of-thousands of GPU threads to securely access data from storage in just microseconds. This is achieved by bypassing the need for a bounce buffer from the host CPU's random access memory (RAM).

The API is accompanied by an open-sourced scaled accelerated data access (SCADA) vertical storage software stack that sits beneath cuFile.

DDN, an early adopter of the platform, is using it to provide more efficient connections between Nvidia GPUs and its Infinia data intelligence platform.