.svg){kind=spacer}
Details
- NVIDIA has introduced RDMA for S3-compatible storage, a technology designed to boost object storage speed for AI training and inference by leveraging Remote Direct Memory Access for data transfers between storage and GPUs, bypassing CPU overhead.
- Industry leaders including Cloudian (HyperStore), Dell Technologies (ObjectScale), and HPE (Alletra Storage MP X10000) have started integrating NVIDIA's RDMA libraries, with broader release slated for January 2026 through the NVIDIA CUDA Toolkit.
- This solution promises higher throughput, reduced latency, lower CPU use, and decreased storage costs per terabyte, offering a direct answer to data transfer bottlenecks that can idle expensive GPU resources during AI workloads.
- With projections anticipating nearly 400 zettabytes of unstructured data annually by 2028, NVIDIA's approach aims to address the limitations of traditional object storage, which was not built for the scale and speed AI workloads now demand.
- NVIDIA's architecture is open, enabling contributions from third-party vendors and users, with the company collaborating on a standardized model for industry-wide adoption of RDMA for S3-compatible storage APIs.
Impact
NVIDIA's RDMA for S3 launch aims to solve a critical challenge in maximizing AI infrastructure investment by eliminating storage bottlenecks. The company's collaboration with major vendors and open architecture push is driving a new industry standard for high-performance data delivery. As competition heats up with other providers exploring similar technologies, NVIDIA cements its leadership in the evolving landscape of AI-centric storage infrastructure.
