NVIDIA Open-Sources cuFile APIs: GPU-Direct Storage Accelerates AI Data Path, Ushering in a New Storage Paradigm
Summary:At the Future of Memory and Storage (FMS) conference, NVIDIA announced the open-sourcing of its cuFile APIs and the underlying vertical storage software stack, enabling GPUs to directly read and write to storage with microsecond-level data access. Additionally, NVIDIA launched the Storage-Next initiative and the SCADA framework, collaborating with over 40 vendors including DDN, KIOXIA, and Micron to drive standardization and innovation in AI storage architecture. The Vera CPU delivers up to 3.21
Core Event
At the Future of Memory and Storage (FMS) conference held in Santa Clara on August 4, 2026, NVIDIA announced a series of major advancements in AI storage, addressing the growing demands of AI workloads on storage infrastructure. The most notable is the open-sourcing of cuFile APIs and the underlying vertical storage software stack, which will fundamentally change the way data is exchanged between GPUs and storage.
Technical Highlights
1. cuFile API Open-Source: GPU-Direct Storage
- Direct Data Path: cuFile APIs allow GPUs to bypass the CPU and directly read/write to storage, leveraging GPU's massive threads and high-bandwidth memory to achieve microsecond-level data access latency.
- Open Source & Collaboration: NVIDIA hosts cuFile APIs on GitHub, with Google, Intel, Meta as initial maintainers, promoting cross-platform optimization.
- Security Foundation: Built on Linux security practices, cuFile enables fast data access for AI-powered defenses, supporting initiatives like the Open Secure AI Alliance.
2. Vera CPU Performance Breakthrough
- Performance Data: In a two-stage compression and encryption pipeline, NVIDIA Vera CPU delivers up to 3.21x higher throughput compared to x86 CPU.
- Significance: This means storage platforms can handle AI data floods more efficiently, reducing compute infrastructure requirements and lowering total cost of ownership.
3. Storage-Next Initiative & SCADA Framework
- Storage-Next: NVIDIA collaborates with over 40 storage and flash vendors including DDN, KIOXIA, Micron to define industry standards for GPU-driven storage, ensuring interoperability.
- SCADA: Stands for Scaled, Accelerated Data Access, a framework that allows GPUs to pull only the necessary data directly into their high-speed memory, avoiding unnecessary data transfers. DDN has integrated SCADA into its Infinia platform.
Industry Impact
- Storage Role Transformation: Storage shifts from a passive data repository to an active part of the data path, redefining the economic tradeoff between memory and storage, with access latency dropping from minutes to microseconds.
- AI Factory Efficiency: By accelerating data access, AI factories can produce accurate, actionable insights faster, improving overall productivity.
- Ecosystem Collaboration: Open-source and standardization efforts promote collaborative innovation across the storage ecosystem, providing developers with more flexible and efficient solutions.
Developer Recommendations
- Adopt cuFile: Developers should evaluate integrating cuFile into existing data pipelines to leverage the performance gains from GPU-direct storage.
- Monitor Storage-Next Standards: Storage vendors and developers should follow Storage-Next progress to ensure compatibility with future standards.
- Security First: While pursuing speed, adhere to SCADA's security model to ensure isolation and protection of data access.
Source
This article is based on NVIDIA's official blog post "As AI Increases Demands on Memory, Storage Steps Up", published on August 4, 2026, available at: https://blogs.nvidia.com/blog/ai-storage-fms/
— END —Tags: #NVIDIA #cuffile #AI STORAGE #GPU DIRECT COMPANY #Open Source AI
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments