NVIDIA Ethernet Adapters PRODUCT LIST

Ethernet adapter selection for data center links

An Ethernet adapter should be selected for the host, traffic pattern and complete link path. Start with the required rate, port count and workload: general server traffic, storage, virtualization and AI communication place different demands on a network interface. Then compare the exact adapter OPN and published capabilities. Depending on the model, 25Gb, 100Gb, 200Gb, 400Gb and 800Gb configurations can require different PCIe interfaces, connectors and cooling. A series name or advertised link rate is insufficient to establish fit in an existing server.

Verify the PCIe generation and lane width at the intended slot, physical clearance, bracket and airflow. Review how the slot connects to the CPU and whether that path supports the application bandwidth target. Check the operating system, driver, firmware and virtualization stack together. If the deployment uses SR-IOV, offloads or other acceleration features, confirm their support for the exact hardware and software combination. Separate what the specification lists from what the application will actually use.

For the physical link, identify the peer switch model, supported port rates, transceiver or DAC/AOC part number, distance and breakout configuration. Match connector, lane mapping and forward error correction requirements at both ends. Similar connector shapes do not establish interoperability. Compare the Ethernet adapter list, Ethernet switches and optical module collection as a single path, while checking each candidate against the manufacturer compatibility guidance.

RoCE adds a fabric configuration requirement beyond basic Ethernet connectivity. Confirm whether the application requires RoCE, then review the supported RDMA software, MTU, queue mapping and congestion control across hosts and switches. PFC and ECN settings depend on the supported design and traffic class; they should not be copied from an unrelated network. A link that comes up successfully can still perform poorly under contention if the fabric policy and application requirements are misaligned.

A useful acceptance plan starts with negotiated speed, error counters and sustained ordinary Ethernet traffic. Add RDMA and congestion tests when RoCE is required, plus the storage, virtualization or training workload that motivated the purchase. Use the planned firmware and driver versions and record the acceptance thresholds. For a quotation, send the exact adapter model, host and switch details, media lengths, quantities, delivery region and schedule. Request a compatibility review or pilot when using mixed generations, unfamiliar optics or a dense chassis, and confirm current availability before approving the purchase.