RDMA

Remote Direct Memory Access lets one node read or write another node's memory over the network without involving either CPU, providing the low-latency, high-bandwidth fabric that multi-node GPU inference clusters rely on to shard a model across machines.

grounded in: Repeated 'spark: 2-node GB10 telemetry refresh' commits (e.g. 384d151, aae33bb) — two DGX Spark GB10 units clustered for multi-node inference are joined over an RDMA-capable interconnect (ConnectX/RoC

Connected concepts

GPU interconnect fabric, Distributed inference, Collective communication, Tensor Parallelism, Zero-Copy, Decentralized mesh inference