InfiniBand overview

InfiniBand (IB) is a high-performance interconnect used in HPC clusters and AI data centres, where latency and throughput between nodes are critical. It is standardised by the InfiniBand Trade Association (IBTA).

What sets it apart

• Very low latency and high bandwidth between servers.
• RDMA (Remote Direct Memory Access) — one node reads or writes another's memory with little CPU involvement.
• Lossless credit-based flow control, like Fibre Channel.

Links and widths

An IB link is built from lanes; common widths are ×1, ×4 and ×12, with ×4 being the usual port. A ×4 port maps naturally onto a QSFP-family module, so InfiniBand and high-rate Ethernet often share the same transceiver housings.

Rates

Each IB generation raises the per-lane signalling rate (SDR, DDR, QDR, FDR, EDR, HDR, NDR …). The per-lane and ×4 aggregate figures are listed in InfiniBand speeds.

Optical InfiniBand modules carry identity and diagnostics in the same managed memory CodingBox reads for other transceivers.

Further reading

• IB protocol & transport — layers, LIDs/GIDs/PKeys, queue pairs and RDMA, credit flow control, the subnet manager, diagnostic tools.
• IB cabling by generation — connectors QDR to XDR, reach classes for DAC/ACC/AOC/transceivers, NDR twin-port and split cables, finned vs flat OSFP.