Network Interface Card (NIC)
Feb 24, 2026| Your server has 100G ports, but your NIC only supports 10G. Congratulations - you just paid for bandwidth you'll never use. The network interface card is the most overlooked bottleneck in modern network design. While engineers obsess over switch fabric and transceiver specs, the NIC quietly determines the actual throughput every server can push. This guide goes beyond basic definitions to explain how NICs work, what types exist, and - most critically - how to match your NIC to the right transceiver, DAC cable, or AOC for maximum performance.
What Is a Network Interface Card?
A network interface card (NIC), also called a network adapter or network interface controller, is the hardware that connects a server, workstation, or storage device to the network. It handles the conversion between data inside the host system and signals on the wire or fiber - managing everything from frame encapsulation and checksum verification to interrupt handling and DMA transfers.
Every NIC has a unique MAC address burned into its firmware, serving as the device's identity on the Layer 2 network. Early NICs shipped as standalone PCI expansion cards. Today, most consumer PCs use a LAN-on-motherboard (LOM) chip for basic 1G connectivity. But in data centers and enterprise environments, dedicated NIC cards remain essential because onboard controllers simply cannot deliver the throughput, offloading capability, or port density that modern workloads demand.
How a NIC Actually Works: Beyond "Send and Receive"
The textbook explanation - "NICs encapsulate data into frames and send them" - is technically correct but misses what makes a NIC valuable. Here's what happens under the hood in a modern server NIC:
Frame Processing
On the transmit side, the NIC takes data from the OS network stack, adds Ethernet headers and CRC checksums, then serializes it into electrical or optical signals. On the receive side, it reverses the process, verifying the CRC and stripping headers before passing the payload to the host.
Offload Engines
This is where NICs earn their keep. Modern NICs offload tasks that would otherwise consume CPU cycles: TCP/UDP checksum computation, TCP segmentation offload (TSO), large receive offload (LRO), and receive side scaling (RSS) that distributes incoming traffic across multiple CPU cores. Without these offloads, a 25G NIC could saturate a CPU core just processing packets.
Kernel Bypass with DPDK
Even with hardware offloads, the OS kernel's network stack introduces overhead at extreme packet rates. NICs that support DPDK (Data Plane Development Kit) allow applications to read and write packets directly from user space, bypassing the kernel entirely. This drops per-packet latency from milliseconds into single-digit microseconds - a requirement for high-frequency trading, 5G user-plane functions, and NFV workloads pushing line-rate on 100G links.
DMA and Interrupt Coalescing
Instead of the CPU copying every packet, the NIC uses Direct Memory Access to write frames directly into system RAM. Interrupt coalescing groups multiple packet events into a single CPU interrupt, dramatically reducing overhead at high packet rates.
Virtualization Support
In virtualized environments, SR-IOV (Single Root I/O Virtualization) lets a single physical NIC present multiple virtual functions, each assigned directly to a VM. This bypasses the hypervisor's virtual switch entirely, delivering near-bare-metal network performance to virtual machines.
NIC Types: A Practical Classification
By Interface Type - This Is What Determines Your Cabling
| NIC Port Type | Connector | Media | Typical Speed | Max Distance |
|---|---|---|---|---|
| RJ45 (Base-T) | RJ45 | Cat5e/Cat6/Cat6a copper | 1G / 2.5G / 5G / 10G | 100m |
| SFP+ | LC duplex | 10G SFP+ transceivers or 10G DAC cables | 10G | 80km (fiber) / 7m (DAC) |
| SFP28 | LC duplex | 25G SFP28 transceivers or 25G DAC | 25G | 40km (fiber) / 5m (DAC) |
| QSFP+ | MPO or LC | 40G QSFP+ transceivers | 40G | 40km (fiber) / 7m (DAC) |
| QSFP28 | MPO or LC | 100G QSFP28 transceivers or 100G DAC cables | 100G | 80km (fiber) / 5m (DAC) |
| QSFP-DD | MPO or LC | 400G QSFP-DD transceivers | 400G | 10km+ (fiber) |
Why this matters: The NIC port type dictates which transceivers and cables you can use. Choosing a dual-port SFP28 NIC locks you into the SFP28 ecosystem - 25G SFP28 optical modules, SFP28 DAC cables, or SFP28 AOC cables. Choose wrong, and you'll be buying adapters or replacing cards.
A note on MPO vs. LC connectors: QSFP+ and QSFP28 ports labeled "SR4" use parallel optics - four lanes of 10G or 25G transmitted simultaneously. This requires MPO/MTP-12 patchcords with 8 or 12 fiber cores, not the standard duplex LC cables used by SFP+ or SFP28 modules. If you're deploying 40G or 100G SR4 optics for the first time, your existing two-fiber patch panel infrastructure will not work - plan for MPO trunk cabling and cassettes.
By Bus Interface - PCIe Generation Matters
All modern server NICs use PCIe. But the PCIe generation and lane width set a hard ceiling on throughput:
- PCIe 3.0 x8: ~64 Gbps - sufficient for dual-port 25G or single-port 40G
- PCIe 3.0 x16: ~128 Gbps - adequate for single-port 100G
- PCIe 4.0 x16: ~256 Gbps - required for dual-port 100G or single-port 200G
- PCIe 5.0 x16: ~512 Gbps - enables 400G NICs
A 100G NIC plugged into a PCIe 3.0 x8 slot physically cannot deliver 100G. Always verify your server's PCIe slot generation and width before purchasing.
By Form Factor - OCP NIC 3.0 Is the New Default
Traditional PCIe add-in cards still dominate enterprise servers, but hyperscale data centers have largely moved to the OCP NIC 3.0 standard (both Small Form Factor and Large Form Factor). OCP NICs plug into a dedicated mezzanine slot on the server motherboard instead of occupying a standard PCIe expansion slot. The benefits are significant: hot-swap capability without powering down the server, improved thermal design with direct airflow over the NIC, and a shorter electrical path to the CPU that reduces signal integrity concerns at 100G+ speeds. If your server platform supports OCP 3.0 - and most modern designs from major ODMs do - it should be your first choice for 100G and 200G NIC deployments.
By Intelligence Level
Standard NICs handle basic packet processing with hardware offloads. Suitable for most enterprise and data center applications.
SmartNICs add a programmable data processing unit (DPU) that can execute firewall rules, encryption, load balancing, and telemetry directly on the NIC, freeing host CPU cores for application workloads. In cloud infrastructure where every CPU core is monetized, SmartNICs pay for themselves by reclaiming cores that would otherwise run OVS or IPsec.
RDMA NICs (RNICs) support Remote Direct Memory Access protocols like RoCEv2 or iWARP, enabling server-to-server memory transfers with single-digit microsecond latency. In GPU training clusters built around NVIDIA H100 or A100 accelerators, RDMA NICs are not optional - they are the backbone. Each GPU node needs 200G or 400G RDMA connectivity to keep the all-reduce gradient synchronization from becoming the training bottleneck. Beyond AI, RDMA NICs are equally critical for HPC workloads and storage systems running NVMe-oF, where kernel-bypass data paths slash storage access latency by 10×.
How to Choose the Right NIC: A Decision Framework
Step 1 - Define your speed requirement. Match it to your switch uplink capacity, not your peak theoretical need. If your ToR switch has 25G server-facing ports, a 25G NIC is the right choice - not 10G (too slow) or 100G (wasted).
Step 2 - Choose your interface type. For in-rack connections under 5 meters, pair your NIC with a DAC cable for the lowest cost and latency. For inter-rack links up to 100 meters, use an AOC cable or a short-reach transceiver with LC fiber patchcord. For building-to-building or campus links, use a long-reach optical transceiver - such as a 100G QSFP28 LR4 module with single-mode fiber and MPO/MTP patchcords.
Step 3 - Verify PCIe compatibility. Check your server's available PCIe slot generation and lane width against the NIC's requirements.
Step 4 - Evaluate offload needs. Running virtualization with dozens of VMs? Prioritize SR-IOV support. AI training cluster? RDMA capability is non-negotiable. Cloud-native microservices? A SmartNIC with programmable pipeline saves CPU cores.
Step 5 - Plan for dual-port redundancy. Production servers should always have at least two NIC ports for link aggregation (LACP) or active-standby failover. Dual-port NICs are more space-efficient than two single-port cards.
Common NIC Mistakes (and How to Avoid Them)
Buying a 10G RJ45 NIC for data center use. 10GBASE-T NICs consume 2–5W more power per port than SFP+ NICs and generate significantly more heat. In a rack with 40 servers, that's 80–200W of unnecessary thermal load. Use 10G SFP+ transceivers with SFP+ NICs instead.
Ignoring firmware and driver updates. NIC firmware bugs cause silent packet drops, CRC errors, and performance degradation. Vendors release critical firmware patches that often aren't applied because the NIC "seems to work fine."
Mismatching NIC speed and switch port speed. A 25G NIC connected to a 10G switch port will auto-negotiate down to 10G - but some NICs handle this poorly, resulting in link flapping or CRC errors. Always verify speed compatibility or use matching equipment.
FAQs
What is the difference between a NIC and a LOM?
A LOM (LAN on Motherboard) is a basic network controller integrated into the server motherboard, typically offering 1G or 10G Base-T connectivity. A dedicated NIC card provides higher speeds (25G–400G), advanced offloads like SR-IOV, and optical interface options for use with fiber transceivers and DAC/AOC cables.
Can I use third-party transceivers in my NIC?
Yes. Most NICs from Intel, Mellanox/NVIDIA, and Broadcom accept MSA-compliant third-party optical transceivers. However, some NIC firmware may display compatibility warnings. Reputable suppliers provide transceivers coded for specific NIC brands to ensure clean operation.
What NIC speed do I need for AI and machine learning workloads?
For GPU training clusters (NVIDIA H100, A100, or similar), each node typically requires 200G or 400G RDMA-capable NICs (RoCEv2) to prevent the network from bottlenecking all-reduce gradient synchronization. Inference serving workloads are less demanding and can often operate on 25G–100G depending on model size and request throughput. Pair training NICs with 400G QSFP-DD modules for spine connections, or use 100G QSFP28 transceivers for leaf-to-server links in smaller clusters.
Is a SmartNIC worth the extra cost?
For cloud providers and large enterprises running heavy network virtualization, yes - SmartNICs can reclaim 4–8 CPU cores per server by offloading virtual switching, security, and telemetry functions. For a standard on-premises deployment with moderate virtualization, a standard NIC with SR-IOV support is usually sufficient.
How do I connect a NIC to fiber optic cabling?
NICs with SFP+, SFP28, QSFP28, or QSFP-DD ports require a matching optical transceiver module to interface with fiber. Insert the transceiver into the NIC port, then connect the appropriate fiber patchcord - LC duplex for SFP-type ports, MPO/MTP for parallel QSFP-type optics. For short-distance links within racks, a DAC cable eliminates the need for both a transceiver and fiber.