InfiniBand provides high-bandwidth, low-latency interconnect between nodes in a GPU cluster for distributed training and multi-node inference workloads.
Gcore configures InfiniBand automatically at cluster creation — all nodes with InfiniBand network cards join the same InfiniBand domain. H100 configurations have eight InfiniBand ports per node, each creating a dedicated network interface.InfiniBand configurations also include SHARP, which offloads collective communication operations to the network switches.
InfiniBand is intended for workloads that require frequent communication between nodes — distributed training and multi-node inference. Workloads that run independently on individual nodes do not use the InfiniBand interconnect.
