NVIDIA open-sources cuFile APIs to advance AI storage standards

At the Future of Memory and Storage (FMS) 2026 conference, NVIDIA announced it is open-sourcing its cuFile APIs, the core component of GPUDirect Storage that enables direct data transfer between GPUs and storage. The code has transitioned to a new GitHub organization, xio-sig, with NVIDIA, Google, Intel, and Meta serving as the inaugural maintainers.

This move aims to address performance bottlenecks in AI workloads, particularly regarding small-block data accesses (512 bytes) that often overwhelm traditional CPUs. To mitigate this, NVIDIA introduced the SCADA (Scaled Accelerated Data Access) framework, which shifts the storage control path from the CPU to the GPU, allowing the GPU to manage its own read requests and completions.

Additionally, NVIDIA is leading the "Storage-Next" initiative, a coalition of over 40 storage and flash vendors—including Micron, KIOXIA, and DDN—to establish interoperable, open standards for GPU-driven storage behavior. Industry partners have demonstrated the potential of these technologies, with Micron showcasing high-throughput performance using multiple SSDs and H100 accelerators.