VDURA adds RDMA & AI-focused context-aware tiering

VDURA has added Remote Direct Memory Access (RDMA) support to its data platform and outlined a new "context-aware tiering" approach for managing data placement across storage layers for AI workloads.

The storage software supplier unveiled the updates around Nvidia's GTC event and also detailed reference configurations pairing its platform with AMD EPYC Turin processors and Nvidia ConnectX-7 networking adapters.

RDMA rollout

RDMA is now available on VDURA V5000 and V7000-class systems running the VDURA Data Platform. The feature enables GPU-to-storage transfers to bypass the CPU by using direct memory access between GPU server nodes and the platform.

In practice, this shifts more data movement onto the network fabric and reduces CPU involvement in moving training and inference datasets between compute and storage. VDURA said all GPU server data transfers now use RDMA across the network tier, aiming to reduce bottlenecks that can arise when CPUs handle data transport.

VDURA linked the RDMA support to a capability it calls DirectFlow, which it described as managing the data path between GPU servers and the platform.

Tiering plans

Alongside the RDMA launch, VDURA previewed the first phase of its Context-Aware Tiering technology, with general availability planned for later this year.

This initial phase includes three functions focused on where hot and persistent data resides during AI training and inference. First, it extends the DirectFlow buffer layer to local NVMe SSDs in the server, reducing reliance on network storage for frequently accessed data and lowering latency for active workloads.

Second is "KVCache writeback," which VDURA described as writing back only persistence-critical KVCache data to durable storage. It said this limits unnecessary I/O activity while maintaining service-level agreements for inference pipelines.

Third is a "Context Cache Tiering" framework, which VDURA said provides read and write access across local SSD and DRAM tiers. It targets inference use cases such as long-context language model serving and retrieval-augmented generation.

VDURA also outlined a broader roadmap through 2027, including deeper application-directed data placement, expanded cross-node cache coherence, and broader hardware support for Nvidia BlueField-4 DPUs.

Hardware combinations

VDURA also outlined infrastructure configurations for its data platform built around AMD's latest EPYC Turin processors and Nvidia ConnectX-7 networking. ConnectX-7 is part of Nvidia's portfolio of high-speed network adapters used in clustered computing environments, including GPU servers for AI training and inference.

By tying RDMA support to specific CPU and networking components, VDURA is aligning its storage stack with hardware commonly used in large AI clusters. It positioned the combination as an optimised option for deployments where network throughput and latency are key constraints.

Focus on AI pipelines

AI operators and infrastructure teams have been working through constraints driven by the volume and velocity of data required for model training, fine-tuning, and serving. Storage can become a limiting factor when GPUs sit idle waiting for data, while cost and complexity rise as organisations add more compute and more storage tiers.

VDURA is framing the RDMA release and its tiering roadmap as responses to those operational issues, while positioning its product as "GPU-native"-reflecting a broader industry push to reduce CPU involvement in moving data for GPU workloads.

CEO Ken Claffey said the updates reflect a broader effort to manage data across memory, flash, and longer-term storage.

"Today's announcements at GTC 2026 reflect our commitment to delivering the AI storage platform that spans the full data hierarchy-from memory to long-term retention-with no compromises on performance," Claffey said. "RDMA gives AI teams direct, zero-CPU-overhead access to their data. Context-Aware Tiering brings intelligence to every tier of the extended storage hierarchy, so data is always in the right place at the right time. Together, these capabilities enable organizations to run larger models, serve more inference requests, and efficiently scale AI infrastructure with the operational reliability that production AI demands."

Context-Aware Tiering Phase 1 is planned for general availability later this year. RDMA support is available now for VDURA V5000 and V7000-class systems running the VDURA Data Platform.