NVIDIA InfiniBand Adapters PRODUCT LIST

InfiniBand adapter selection for GPU and HPC servers

An InfiniBand adapter connects a server to an InfiniBand fabric for workloads such as HPC, distributed storage and GPU cluster communication. Selection starts with the application and fabric requirements, then moves to the host platform. Identify the intended HDR or NDR rate, number of ports and switch generation. Keep the full ConnectX model and OPN in the shortlist: products within one family can differ in port mode, connector, host interface and supported features. Do not infer the required configuration from the family name alone.

Check the server before comparing nominal link speeds. Confirm the PCIe generation and lane width available at the actual slot, CPU attachment, bracket dimensions, chassis clearance and cooling requirements. The adapter's published network rate is not a promise of application throughput. A constrained PCIe slot or an unsuitable CPU and memory path can become the limiting factor. For GPU systems, review the supported GPU, driver and host topology when direct communication features are required, and verify support in the manufacturer's documentation.

The fabric connection must be checked end to end. Match the adapter port to the switch port using a supported cable or optic with the correct connector, lane mapping, length and rate. Review breakout support at both endpoints and confirm whether a VPI adapter will operate in InfiniBand mode. Port mode is a model-specific capability; an Ethernet-only adapter should not be selected for an InfiniBand fabric. Compare the adapter models with the InfiniBand switches and link media before freezing the bill of materials.

Record the operating system, kernel, driver or OFED branch and firmware version for every host group. Check the supported combination rather than upgrading one component in isolation. The deployment plan also needs a subnet manager, addressing and partition requirements where applicable, and a repeatable method for discovering and monitoring the fabric. Maintain a rollback path for firmware or driver changes on existing production nodes.

Validate a candidate on a representative server and peer switch with the planned software image. Begin with link state, negotiated rate and error counters, then test RDMA communication, application throughput, latency and sustained stability. GPU training tests should reflect the intended node count and communication pattern as far as the pilot allows. Include quantities, spare policy, delivery date and acceptance criteria in the inquiry. This provides a concrete basis for model selection, compatibility review and rollout planning without treating a laboratory result as a guarantee for every deployment.