NVIDIA announced that xio-sig will be expanded to include NVIDIA cuObject alongside cuFile, and that cuObject client and server libraries are now generally available. The company also released a Scaled Accelerated Data Access (SCADA) Server SDK to let storage providers build servers that respond to GPU-initiated requests and deliver data over RDMA.
Why this matters for AI infrastructure
AI workloads — such as training, fine-tuning, inference context, tool calls, searches and database lookups — increasingly demand high-speed access to large datasets. Much of that data is stored as files and objects either on-premises or in the cloud. Compute accelerators (GPUs, TPUs, XPUs) benefit from remote direct memory access (RDMA) that uses NIC- or DPU-accelerated transfers (for example NVIDIA ConnectX NIC or NVIDIA BlueField DPU) and avoids copying data through server CPU-controlled memory. As GPU architectures become faster, the need for RDMA-accelerated, zero-copy transfers grows.
Until now, developers implementing direct access to file and object storage had to deal with multiple APIs and provider-specific protocols. Object storage over RDMA lacked a common wire protocol, often forcing provider-specific integrations or fallback to traditional access methods.
cuObject in xio-sig and library availability
The general availability of cuObject client and server libraries gives AI application developers, open-source framework maintainers, storage providers and consumers a standardized way to accelerate data access for both file and object protocols. Using cuObject APIs and an RDMA wire protocol, developers can build accelerated object-storage applications and servers such that object data transfers occur over RDMA without routing data through server CPUs. This enables higher throughput, lower latency and reduced CPU utilization for reads and writes.
The xio-sig repository layout now separates cuFile and cuObject. Headers for cuFile and cuObject, the cuObject wire protocol, and implementation code for libxFile and xFilekernel will be shared once the production-ready stack passes conformance tests. Governance documents are under review by pending Board members.
Google Cloud, previously involved as a maintainer for cuFile, is evaluating broader participation for cuObject, reflecting its focus on high-performance cloud file and object storage. Microsoft has also signaled interest in joining the xio-sig Board to improve interoperability in storage I/O.
SCADA Server SDK and the Storage-Next initiative
As part of the Storage-Next effort, the new SCADA Server SDK lets storage partners build servers that receive requests from GPU-based SCADA clients, fulfill them from local or remote storage, and return results over RDMA. The work also includes a Storage Lender Service and a SCADA command-line utility for configuration and deployment, designed to help storage-provider servers interoperate with SCADA clients.
IBM Storage has demonstrated a prototype in which a SCADA client sent requests to an initial Storage Scale SCADA server built with the SCADA Server SDK. That demonstration shows how storage vendors can collaborate with NVIDIA to build an ecosystem supporting accelerated, GPU-initiated storage access, which could enable better access to large datasets for semantic search, recommender systems and fraud detection.
Industry collaboration and next steps
Through Storage-Next, NVIDIA is leading a consortium of more than 40 vendors and customers — including NAND vendors, controller vendors, storage providers, hyperscalers and application developers — to define GPU-driven storage behavior and convert those advances into interoperable, open industry standards. SCADA forms the software infrastructure for high-throughput, fine-grained, GPU-initiated I/O.
With cuObject libraries generally available and xio-sig expanded, storage partners, providers and consumers can begin adopting cuObject and contribute to the community work on interoperable APIs and protocols for cuFile and cuObject. Relevant resources include the xio-sig repository updates, the NVIDIA FMS blog posts on cuFile and Storage-Next, NVIDIA's SCADA presentations (including at the 2024 OCP Summit), and open-source tooling and documentation such as the Storage Lender Service and GPUDirect Storage (GDS) documentation.
Developers and storage providers should watch the xio-sig repository, try cuObject and the SCADA Server SDK in their environments, and prepare systems for GPU-accelerated, RDMA-based storage access.
