Nvidia Open-Sources CuFile Storage Stack, Sets Industry Standard for GPU Storage

Nvidia has open-sourced its cuFile storage stack, making it an industry standard for GPU storage. The move is backed by Google, Intel, and Meta, who have joined Nvidia as founding maintainers of the Accelerated IO Special Interest Group on GitHub.

CuFile allows a GPU to read from and write to storage directly, without needing the CPU to fetch data first. This reduces latency and increases performance, making it ideal for applications that require high-speed data transfer, such as AI and machine learning workloads.

Nvidia has been working on cuFile for years, with its GPUDirect Storage technology shipping in 2021. However, this new version of the software, dubbed SCADA (scaled accelerated data access), moves the control path onto the GPU, allowing it to build its own storage requests and handle completions.

The result is a significant increase in performance, with Micron reporting a server capable of 230 million random reads per second at 512 bytes each. This is roughly 95% of what those drives are rated to deliver, and would have required dozens of CPU cores to achieve using host software.

Nvidia's stated reason for open-sourcing cuFile is security and Linux interoperability. By making the code open-source, developers can ensure that applications can talk to drives directly without needing to go through a proprietary layer, reducing the risk of data breaches and improving overall system reliability.