A network interface card (NIC) is the hardware component that connects a computer or server to a network, converting data between the system’s memory and the electrical, optical or radio signals of the link. In network hardware, the NIC implements the physical and data link layers: it owns a MAC address, builds and checks Ethernet frames, and moves packets to and from the host over the PCI Express bus.
The name dates from the era of separate expansion cards, but the function is the same whether it sits on a plug-in card, on the motherboard, in a USB adapter or in a smart network device with its own processors. Servers, desktops, laptops and network appliances all depend on one or more NICs.
How a network interface card moves data
A NIC moves data through a pipeline of three main parts: the PHY, which handles signaling on the cable or fiber, the MAC controller, which builds and validates frames, and a DMA engine, which copies packets directly between the card and system memory without involving the CPU for every byte.
The driver and the card share descriptor rings in memory. For transmission, the driver places packet addresses in a transmit ring and the card fetches the data; for reception, the card writes incoming frames into preallocated buffers and raises an interrupt. At high rates, interrupt moderation and polling, as in Linux NAPI, batch this work to avoid overwhelming the CPU.
| Component | Function | Typical detail |
|---|---|---|
| PHY or transceiver | Encodes bits onto the medium | RJ45 copper or SFP, SFP+ and QSFP cages |
| MAC controller | Frames, addresses and checks CRC | Holds the burned-in 48-bit MAC address |
| DMA engine | Transfers data to host memory | Uses descriptor rings in RAM |
| Host interface | Connects to the system bus | PCIe 3.0 to 5.0, x4 to x16 lanes |
| Firmware and offload logic | Performs checksum and segmentation work | Updated with vendor tools |
Speeds and physical interfaces
NIC speed is set by the Ethernet standard and the connector type, and the PCIe connection must provide enough bandwidth to keep up. A dual-port 100 Gbit/s card, for example, needs at least a PCIe 4.0 x16 slot to run both ports at full rate.
| Speed | Typical connector | Usual setting |
|---|---|---|
| 1 Gbit/s | RJ45 over CAT5e or CAT6 | Desktops, laptops, management ports |
| 2.5 and 5 Gbit/s | RJ45 over CAT5e or CAT6 | Workstations, Wi-Fi access point uplinks |
| 10 Gbit/s | RJ45 10GBASE-T or SFP+ | Servers, storage, small office backbones |
| 25 Gbit/s | SFP28 | Data center server access |
| 100 Gbit/s and above | QSFP28, QSFP56 or QSFP-DD | Hyperscale, storage and AI clusters |
SFP-based cards offer flexibility because the transceiver or direct attach cable determines the medium, but optics must be compatible with the card. Some vendors lock cards to approved modules, which is a frequent cause of links that simply stay down.
Offloads and server NIC features
Server NICs differ from basic adapters mainly through offload features that move repetitive work from the CPU to the card. These features determine how many packets per second a host can process and how much CPU remains for applications.
| Feature | What it does | Benefit |
|---|---|---|
| Checksum offload | Computes IP, TCP and UDP checksums | Saves CPU cycles per packet |
| TCP segmentation offload | Splits large buffers into MSS-sized segments | Fewer, larger transmit operations |
| Receive side scaling | Spreads flows across multiple queues and cores | Scales receive processing |
| SR-IOV | Presents virtual functions to virtual machines | Near-native performance for guests |
| RDMA over Converged Ethernet | Transfers memory directly between hosts | Very low latency for storage and clusters |
| Wake-on-LAN and PXE | Starts or boots a system over the network | Remote management and provisioning |
At the top end, SmartNICs and DPUs add programmable processors or FPGAs that run virtual switching, encryption or storage protocols on the card itself. They suit cloud and high-density environments but bring new firmware and management responsibilities.
Wireless NICs follow the same architecture with a radio in place of the Ethernet PHY. They implement IEEE 802.11 framing, encryption and power saving, and their real throughput depends on antenna count, channel width and signal quality far more than on the nominal rate printed on the box. In laptops they are usually small M.2 modules, which makes upgrading to a newer Wi-Fi generation possible when the firmware and operating system allow it.
Every NIC also carries an identity. The burned-in MAC address is assigned by the manufacturer from its registered address block, but drivers can override it, and modern operating systems often use randomized MAC addresses on Wi-Fi to limit tracking. Network access control and DHCP reservations that rely on MAC addresses must account for this behavior.
Choosing and troubleshooting a NIC
Choosing a NIC means matching speed, connector, host bus and driver support to the workload and the switch it connects to. Driver maturity in the operating system or hypervisor often matters more than raw specifications, especially for virtualization hosts.
Most NIC problems fall into a few patterns. The table lists the ones that administrators meet repeatedly.
| Symptom | Likely cause | Fix |
|---|---|---|
| Link negotiates at 100 Mbit/s | Damaged cable or a wire pair missing | Replace the cable and check the patch panel |
| Rising CRC errors | Faulty cable, optic or duplex mismatch | Swap components and set both ends to auto-negotiate |
| Link stays down with an SFP | Incompatible or wrong-type transceiver | Use a supported module of matching wavelength |
| Throughput far below link speed | Card in a slot with too few PCIe lanes | Move the card to a suitable slot |
| Packet drops under load | Receive rings too small or one queue saturated | Enlarge rings and enable RSS |
For redundancy and bandwidth, servers often combine two ports with link aggregation (LACP) or operating system teaming. A single flow still uses only one member link, so aggregation raises total capacity rather than the speed of one transfer.
Explained without jargon
A network interface card is the doorway between a computer and a network. Everything the computer sends or receives passes through it, whether over a cable, a fiber or the air.
Think of it as a translator at a border. Inside, the computer speaks the language of its memory and processor. Outside, the network speaks in pulses of electricity or light. The card translates in both directions and does it fast enough for video, files and web pages to flow smoothly.
A typical use case
A video production studio stores its footage on a shared file server reached over 1 Gbit/s Ethernet. Editors working with high-resolution footage see constant pauses, because each workstation can read at most about 115 MB/s from the server.
The studio installs 10 Gbit/s SFP+ NICs in the server and in four editing workstations and connects them to a switch with matching ports through direct attach cables. Throughput rises to roughly 1 GB/s per workstation, limited now by the server’s disk array rather than the network. The editors can scrub through footage directly on the server instead of copying files to local disks first.