NVIDIA Open-Sources cuFile APIs: Accelerating AI Storage Performance and Security
Summary:At the 2026 Future of Memory and Storage (FMS) conference, NVIDIA announced the open-sourcing of its cuFile application programming interfaces (APIs), enabling GPUs to directly read from and write to storage, thereby enhancing AI data processing efficiency. As part of NVIDIA GPUDirect Storage, cuFile optimizes storage architecture and data services performance, significantly boosting AI storage system throughput and security. Additionally, NVIDIA introduced the Storage-Next initiative, collabora
NVIDIA Open-Sources cuFile APIs: A New Breakthrough in AI Storage Performance and Security
At the 2026 Future of Memory and Storage (FMS) conference, NVIDIA announced a significant move: the open-sourcing of its cuFile application programming interfaces (APIs). This initiative aims to address the growing demands of AI on storage systems and enhance data processing efficiency. Here are the key highlights of this release:
1. Direct GPU Access to Storage
The cuFile API enables GPUs to directly read from and write to storage, eliminating the need for CPU mediation. This leverages the high parallel computing power of GPUs, significantly reducing data transfer latency.
2. Enhanced AI Storage Performance
As part of NVIDIA GPUDirect Storage, cuFile optimizes storage architecture and data services performance, achieving:
- High Throughput: The NVIDIA Vera CPU delivers up to 3.21x higher throughput than an x86 CPU in a two-stage compression and encryption pipeline.
- Low Latency: Using GPU threads and high-bandwidth memory, cuFile can securely access storage data in microseconds.
3. Strengthened AI Storage Security
cuFile emphasizes not only performance but also security. Its design is based on Linux best practices, providing a security-first storage stack that supports AI-driven security defenses, such as real-time data encryption, integrity checking, and rapid recovery.
4. Storage-Next Initiative
NVIDIA also introduced the Storage-Next initiative, collaborating with over 40 leading storage and flash vendors, including DDN, KIOXIA, and Micron, to drive the standardization and innovation of AI storage technologies. The core of this initiative is the SCADA (Scaled, Accelerated Data Access) framework, which allows massively parallel GPUs to pull only the necessary data directly from storage into their own high-speed memory, improving data access efficiency for AI applications.
5. Industry Impact and Future Outlook
NVIDIA's cuFile open-sourcing and Storage-Next initiative mark a significant transformation in the AI storage landscape. Through collaboration with industry leaders, NVIDIA aims to build an efficient, secure, and standardized AI storage ecosystem, providing stronger infrastructure support for AI applications.
Technical Highlights
- GPUDirect Storage: By enabling direct GPU access to storage, cuFile eliminates bottlenecks in traditional storage architectures.
- High-Performance Data Services: The NVIDIA Vera CPU excels in data encryption, compression, and integrity checking.
- Open Collaboration: The open-source nature of cuFile and the cross-industry collaboration of Storage-Next foster rapid development in AI storage technology.
Industry Impact and Developer Recommendations
- AI Application Developers: Consider integrating cuFile APIs into AI applications to enhance data processing efficiency.
- Storage System Vendors: Actively participate in the Storage-Next initiative and collaborate with NVIDIA to drive the standardization of AI storage technologies.
- Enterprise Users: Evaluate existing storage architectures and assess the potential benefits of NVIDIA AI storage solutions.
Original Source
— END —Tags: #NVIDIA #AI Storage #cuFile #GPUDirect Storage #Storage-Next
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