NVIDIA makes cuObject generally available
NVIDIA has released its cuObject client and server libraries and a new SCADA Server SDK, and is folding cuObject into the xio-sig group with Google Cloud and Microsoft to standardize GPU-to-storage RDMA access.
NVIDIA announced general availability of its cuObject client and server libraries, which let GPUs and other accelerators read and write object storage over RDMA. The data skips the server CPU's memory, and NVIDIA says this raises throughput, cuts latency and lowers CPU use. The company also launched a SCADA Server SDK so storage providers can build servers that answer storage requests initiated by GPUs.
NVIDIA's pitch is that storage access for AI is fragmented today. Developers work with different APIs and protocols from each provider, and object storage over RDMA has had no common wire protocol.
To address that, NVIDIA is expanding xio-sig, the group that already covers its cuFile API, to include cuObject and its RDMA wire protocol. Google Cloud and Microsoft are partners. The stated goal is to let the cuObject client work with any server that follows the protocol. Google Cloud, already a cuFile maintainer in the group, is "evaluating" a similar role for cuObject. Microsoft "looks forward to" joining the xio-sig Board. The group's governance documents are still being reviewed by prospective board members. Headers, the wire protocol spec and library code will be published once the production-ready stack passes conformance tests.
The SCADA SDK is part of NVIDIA's Storage-Next initiative, which the company says brings together more than 40 vendors and customers to define open standards for small, fine-grained I/O requests initiated by GPUs. The SDK comes with a Storage Lender Service and a command-line utility for configuring and deploying SCADA. IBM has shown a prototype SCADA server built on IBM Storage Scale.
NVIDIA did not publish performance figures or a timeline for the conformance-tested open-source release.