How RDMA Became the Fuel for Fast Networks (2020) (blogs.nvidia.com)

What started as an obscure idea codified in a 1993 HP patent—allowing one machine to read/write another machine’s RAM without involving the remote CPU or OS—became the backbone of modern high-performance networking. Two chance encounters in the early 2000s pushed RDMA into the mainstream: Ohio State’s D.K. Panda ported MPI stacks to Mellanox’s InfiniBand cards (MVAPICH), and a daring Virginia Tech project used InfiniBand to link 1,100 Power Mac G5s into a 10.3‑teraflop system that landed on the TOP500. Those wins proved RDMA’s latency and throughput advantages over proprietary fabrics and helped Mellanox scale from an HPC startup into a pervasive networking supplier. Technically, RDMA slashes CPU overhead and round‑trip messaging by enabling zero‑copy transfers and direct memory access across nodes—ideal for latency‑sensitive HPC and data‑intensive AI. InfiniBand’s success led to RDMA over Converged Ethernet (RoCE) standards (initial in 2010, routed in 2014) and Mellanox’s ConnectX silicon, which brought 25/50/100G Ethernet support to hyperscalers. Today RDMA is often embedded in smart NICs that offload routing, congestion control, security and collective operations (e.g., Mellanox SHARP) to accelerate distributed training and storage. The technology’s fusion with GPUs and Nvidia’s acquisition of Mellanox underscores RDMA’s central role in scaling next‑generation AI and HPC infrastructure.