NVIDIA Releases cuObject and SCADA SDK for GPU-Accelerated AI Storage

NVIDIA has officially released the general availability of its cuObject client and server libraries, alongside the new SCADA Server SDK, marking a significant step toward standardized, high-performance storage access for artificial intelligence workloads. The release addresses the growing demand for rapid, secure data retrieval as AI models increasingly rely on vast files and objects stored across on-premises and cloud environments. Traditional data transfer methods routing through server CPUs create bottlenecks, whereas NVIDIA’s new tools enable remote direct memory access via NVIDIA ConnectX NICs and BlueField DPUs. This architecture facilitates zero-copy data transfers directly between accelerators and storage, significantly reducing latency, lowering CPU utilization, and increasing throughput for training, fine-tuning, and inference tasks. Central to this effort is the expansion of the xio-sig storage interoperability initiative to include NVIDIA cuObject alongside the existing cuFile standard. By introducing a common RDMA wire protocol, xio-sig allows developers to build accelerated object-storage applications that remain interoperable across different server-side implementations. Google Cloud is currently evaluating expanded participation in the initiative, while Microsoft has signaled its intent to join the xio-sig Board to further improve storage I/O interoperability. Governance documents and repository structures are currently under review to ensure production-ready conformance testing before open-source code is fully shared. Complementing cuObject, the newly available SCADA Server SDK empowers storage providers to develop servers that respond directly to GPU-initiated requests. IBM has already demonstrated interoperability between a SCADA client and a prototype server built on the SDK, integrating it with IBM Storage Scale. This collaboration highlights how third-party vendors can leverage NVIDIA’s infrastructure to support fine-grained, high-throughput I/O operations that do not align with traditional storage media. The broader SCADA software stack also includes a Storage Lender Service and a configuration utility to streamline deployment across distributed environments. These announcements are closely tied to NVIDIA’s wider Storage-Next initiative, which unites over forty vendors, hyperscalers, and developers to establish open industry standards for GPU-driven storage. The project specifically targets the emerging requirement for small, accelerator-initiated requests, aiming to replace fragmented protocols with unified, scalable architectures. The resulting software infrastructure is designed to power next-generation applications such as semantic search, real-time recommender systems, and enterprise fraud detection. With the cuObject libraries and SCADA SDK now accessible, AI infrastructure engineers, storage developers, and cloud service providers can begin integrating accelerated data pathways while preparing to contribute to the evolving ecosystem of interoperable APIs and protocols.