NVIDIA Launches cuObject to Speed Up GPU Storage Access

NVIDIA has announced the general availability of its cuObject client and server libraries, alongside the release of cuObject Server 2.0.0. This technology introduces standardized APIs and a Remote Direct Memory Access (RDMA) wire protocol designed to accelerate object storage access. By utilizing RDMA, the system transfers data directly to compute accelerators like GPUs, TPUs, and XPUs without routing it through the host server's CPU-controlled memory. This direct path reduces latency, increases throughput, and lowers CPU overhead during intensive data reads and writes.

To drive broader adoption and interoperability, the xio-sig consortium is expanding its scope to include cuObject alongside the existing cuFile technology. Tech giants are already backing the initiative, with Microsoft planning to join the xio-sig board and Google Cloud currently evaluating its participation. This collaborative effort aims to establish a unified framework where cuObject clients can seamlessly interoperate with any storage server that adheres to the established wire protocol.

Alongside these updates, NVIDIA introduced the Scaled Accelerated Data Access (SCADA) Server SDK. This toolkit allows storage providers to build servers capable of responding directly to GPU-initiated requests from SCADA clients. IBM Storage has already demonstrated the practical viability of this system, showcasing a prototype that integrates SCADA with its IBM Storage Scale platform. The software infrastructure also includes a Storage Lender Service and a dedicated command-line utility to simplify configuration and deployment.

These releases fit into the broader NVIDIA Storage-Next initiative, a collaborative group of over 40 vendors and customers. This coalition, which includes NAND and controller vendors, hyperscalers, and application developers, is working to define open industry standards for fine-grained, GPU-driven storage access. For AI practitioners, these advancements mean more efficient pipelines for data-heavy tasks like semantic search, recommender systems, and fraud detection, freeing up valuable CPU cycles for other processing needs.