InfiniBand
InfiniBand is a high-bandwidth, low-latency networking technology developed by Mellanox (now NVIDIA Networking) that serves as the primary interconnect fabric for GPU training clusters. InfiniBand delivers up to 400 Gb/s per port (NDR generation) with sub-microsecond latency, enabling the all-reduce communication patterns essential for distributed training. NVIDIA controls the InfiniBand supply chain end-to-end — switches, cables, host channel adapters, and the software stack — making it both the performance leader and a significant bottleneck for cluster procurement.
InfiniBand uses a switched fabric topology, typically deployed in fat-tree or Clos configurations. Each NVIDIA DGX or HGX system connects via multiple InfiniBand ports, creating a non-blocking fabric that allows any GPU to communicate with any other GPU at full bandwidth. The ConnectX-7 adapters used in current deployments support NDR (400G), with XDR (800G) shipping in 2026. InfiniBand also supports RDMA (Remote Direct Memory Access), bypassing the CPU for GPU-to-GPU data transfers.
InfiniBand availability is a critical variable in cluster procurement timelines. Operators who cannot secure InfiniBand allocation face 6-12 month delays. In our deployment advisory work, we assess whether operators have secured sufficient InfiniBand capacity to match their GPU orders — a mismatch that has derailed multiple deployments we have reviewed.
This glossary is maintained by Disintermediate as a reference for GPU infrastructure professionals, investors, and operators. Each entry reflects terminology as used in active advisory engagements and market intelligence work.