How to choose a server network adapter
Size for the fabric you are growing into, not the one you have: 10/25 GbE for general nodes, 100/200 GbE with RoCE v2 for AI and HPC east-west traffic. ConnectX-5 and later speak RoCE v2; Intel E810 speaks both iWARP and RoCE — pick the family your switches and storage stack already run, and check SR-IOV support if you virtualise.
Match form factor and connector: OCP 3.0 SFF and standard PCIe cards are not mechanically interchangeable, so confirm the cage or riser against your chassis. QSFP56 ports accept QSFP28 optics at reduced rate — useful when migrating switch generations.
A 400G NDR card needs a PCIe 5.0 x16 host slot to reach line rate; in a PCIe 4.0 slot it trains down. The 200G dual-port variants remain the workhorse for mixed virtualisation and storage fabrics.
Budget the optics with the card — transceivers, AOC or DAC are the weak link in most bundles. We kit matched, coded optics and pre-verify them with the card at rated speed; specify reach and connector types in your BOM and they arrive plug-compatible.
Tell us the workload, switch generation and host platform — we will spec the exact variant, optics and cabling before quoting, and confirm the offload licence state on used stock.
ConnectX-7 vs Intel E810 — choosing server NICs for AI clusters and RoCEv2 fabrics