NVIDIA GPUDirect RDMA for S3

NVIDIA's RDMA client/server library stack for S3-compatible object storage, introduced November 2025, that moves data directly from a storage node's memory to GPU high-bandwidth memory over RoCE v2 or InfiniBand fabrics — bypassing the host OS kernel, the TCP/IP stack, and user-space copy chains. A storage-side server library integrates into object-storage controllers (MinIO AIStor, Cloudian HyperStore, Dell ObjectScale, HPE Alletra Storage MP X10000); the client library runs on GPU compute nodes or is offloaded to a BlueField-3 DPU using the ROS2 (RDMA-first Object Storage) pattern.

At 400GbE and above, TCP interrupt handling and kernel-space packet copies starve GPUs of training data. The host CPU cannot feed PCIe fast enough, GPUs idle, and the economics of the AI factory collapse. Kernel-bypass RDMA shifts the entire data-transport burden off the CPU, keeping expensive GPUs compute-bound during massive checkpoint loads and distributed training reads.

Multi-trillion-parameter LLM pre-training, synchronized checkpoint loading across thousands of GPU nodes, high-throughput RAG vector ingestion at training scale, SmartNIC-offloaded object storage on BlueField-3 DPUs, on-prem AI factories where standard HTTP S3 is the throughput-per-watt bottleneck.