NVIDIA open-sources cuFile to accelerate AI storage performance

At the Future of Memory and Storage (FMS) conference in Santa Clara, NVIDIA announced the release of its cuFile API and associated storage software stack as open source. This move aims to standardize GPU-to-SSD communication and reduce reliance on closed ecosystems.

Previously, cuFile functioned as a proprietary interface within NVIDIA's GPUDirect Storage technology, allowing data to move directly between NVMe drives and GPU memory via DMA, bypassing the CPU. To manage the open-source transition, a new GitHub organization named 'xio-sig' has been established, with founding maintainers including NVIDIA, Google, Intel, and Meta.

A key technical advancement introduced is the SCADA (Scaled Accelerated Data Access) framework. This framework addresses inefficiencies in AI workloads, specifically the high overhead caused by small 512-byte data accesses that traditional 4KB-optimized SSD controllers struggle to manage. SCADA allows the GPU to handle its own read requests and control path, significantly reducing CPU load and latency.

NVIDIA is collaborating with over 40 manufacturers, such as Micron and Kioxia, through its Storage-Next initiative to integrate these technologies across the industry.