What Is a Network Interface Card? Types, Speeds and Ports

A network interface card (NIC) is the hardware that connects a computer to a network. That sentence hides a lot of engineering. The NIC terminates the physical medium — copper pairs or a fiber — speaks the link protocol, moves frames between the wire and the host’s memory over the PCIe bus, and increasingly performs work the host CPU would rather not do: checksum and segmentation offloads, virtualisation switching, tunnel encapsulation, encryption. Every server in every data centre has at least one; choosing it well is a genuine design decision, as our server network card range illustrates with its nearly one hundred specifications.

The anatomy, briefly

Four parts define a NIC. The controller is its brain — the chip that implements the Ethernet MAC and the offload features; on some of our adapters the controller is our own silicon, such as the G810-2Q. The host interface connects the card to the computer, almost always PCI-Express today, in widths from x4 to x16. The ports face the network: fixed copper RJ45, fixed optical, or empty cages that accept pluggable modules. And the bracket and form factor decide which chassis the card physically fits — including the mezzanine variant defined by the OCP specification, which abandons the PCIe slot for a dedicated motherboard slot.

Types you will meet

• Copper NICs: RJ45 ports, 1G/10G over structured cabling. The default for endpoints and short runs.
• Fiber NICs: optical ports or SFP/SFP+/SFP28/QSFP28 cages. The data-centre standard, for distance and isolation.
• OCP mezzanine cards: for OCP-style servers, plugging into the dedicated slot — see the F2502E-OCP-V3.0.
• Bypass NICs: port pairs with fail-safe bypass hardware for inline appliances.
• Smart/accelerated NICs: heavy offload, RDMA, FPGA or encryption functions on the card itself — the territory of our encryption adapters and custom FPGA practice.
• USB adapters: the portable edge of the family, including security endpoints like the TYPEC-JM-T encryption adapter.

Speed tiers and what they are for

Gigabit (1G) remains the utility tier — management, endpoints, legacy. 10G was the data-centre default for a decade and still matches most installed fabrics. 25G is the current server standard in new fabrics, and 40G/50G occupy transitional niches. 100G serves storage, HPC/AI and aggregation-heavy nodes. The rate on the box is a ceiling, not a promise: the link runs at the slower of the card, the switch port and the medium — which is why our selection guide starts from the fabric and works backwards to the card.

Ports: count and shape

Single-port cards serve simple endpoints. Dual-port is the server default — redundancy, bonding, or a bypass pair — and quad-port serves dense or multi-network nodes. Port shape matters as much as count: two ports on one card share one PCIe connection and one failure domain, which is either an economy or a risk depending on the design. Redundancy purists put the two paths on two different cards; pragmatists accept one card and two switches. Both are defensible; be deliberate about which you chose.

Reading a NIC specification like an engineer

Five lines of any spec sheet carry the decision: controller (whose silicon, which offloads), host interface (slot width and generation — the real bandwidth ceiling), ports (count, cage or fixed, media), features (SR-IOV, RDMA, offloads, bypass), and the platform list (operating systems and, where relevant, CPU adaptation). If a listing hides any of the five, treat the price as provisional. Our product pages publish all five for the flagship models — for example the F1102EBPLR page documents everything from its 256 virtual channels to its 24.5 W power draw — because buyers should not have to ask.

A short history that explains the present

The NIC began as exactly what the name says — a card, slotted into a bus, doing little more than moving frames between wire and memory while the CPU did the thinking. Three forces changed it. Speed: at gigabit and beyond, per-packet CPU work became the bottleneck, so checksums, segmentation and eventually whole protocol functions migrated onto the card. Virtualisation: one physical server became dozens of tenants, so the card learned to present itself as many virtual cards (SR-IOV) and to switch between them in silicon. And specialisation: security appliances needed bypass hardware, storage needed RDMA, regulated industries needed encryption in dedicated silicon — each pulling a NIC sub-family into existence. The modern high-end NIC is therefore less an interface than a specialised co-processor that also faces the network. Understanding that trajectory makes datasheets readable: every feature on a flagship card is a scar from one of those three forces.

Form factors you will actually meet

The PCIe card is the canonical NIC, but the family wears several shapes. Low-profile and full-height PCIe cards differ only in bracket and board outline, yet the difference decides fit in half the world’s 1U and 2U servers — always order the bracket by the chassis, never by habit. OCP mezzanine cards abandon the slot for a motherboard-standardised home in OCP-style servers, trading universality for density and serviceability. USB adapters carry the function to the edge — laptops, field kits, and security endpoints like our encryption adapter. And at the far end of integration, the NIC dissolves into the motherboard itself as a LAN-on-motherboard controller, which is why some servers "have no NIC card" while networking perfectly well. Each shape answers the same question — where does the network function live? — for a different mechanical world.

So the next time a server’s networking misbehaves, resist reaching first for the cable. Ask which member of this family you are dealing with, which of its offload and virtualisation features the workload is actually using, and whether the driver generation matches the operating system’s. The NIC stopped being a passive socket a long time ago; troubleshooting it like one is the anachronism, not the card.