NVIDIA opens new paths for GPU-driven access to AI storage

NVIDIA has made its cuObject client and server libraries generally available and introduced a software kit for storage servers handling GPU-initiated requests. The push could make it easier to move AI data without routing transfers through the server CPU, while work on shared protocols aims to reduce the current patchwork of integrations.

NVIDIA Watch analysis

What happened

The cuObject libraries provide APIs and an RDMA wire protocol for accelerated object-storage access. NVIDIA says the xio-sig consortium is expanding to include cuObject alongside cuFile, its file-storage library. Google Cloud is evaluating participation, while Microsoft plans to join the consortium’s board.

A separate SCADA Server SDK lets storage providers build servers that respond to requests from GPUs through SCADA clients. IBM Storage has demonstrated a prototype integrating SCADA with Storage Scale. NVIDIA describes Storage-Next, an initiative involving more than 40 vendors and customers, as a forum for developing open standards for GPU-driven storage. The NVIDIA announcement sets out the APIs and projects involved.

Why it matters

AI systems can spend plenty of time waiting for data, not just processing it. Direct, RDMA-based transfers could improve throughput, lower latency and reduce CPU use, according to NVIDIA. Shared APIs and protocols could also spare developers and storage providers from building a separate integration for every combination of accelerator and storage system.

That is the opportunity, not a guarantee that today’s storage estates will suddenly work together. The IBM example is a prototype, and NVIDIA says parts of the cuObject production-ready stack will be shared after conformance testing. The practical test is whether other providers can implement the interfaces and connect reliably.

Our read

This is the unglamorous infrastructure work that can decide whether a large AI system spends its time computing or waiting for its files to arrive. NVIDIA is offering both a usable library and a longer-term standards push; the latter will matter only if other storage vendors can join on workable terms. Engineers should treat cuObject as ready for evaluation, while treating the broader interoperability picture as still under construction.

What to watch

• Whether other storage providers build servers or clients compatible with cuObject and SCADA.
• What the xio-sig conformance tests require and when the production-ready code becomes available.
• Whether Google Cloud joins the cuObject work and Microsoft takes its planned board seat.

Discussion spark: For AI storage interoperability, should providers rally around shared APIs and protocols, or is performance better served by tightly integrated vendor-specific systems?

Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.