US How a SmartNIC works
A SmartNIC, or smart network interface card, is a network adapter that carries its own processing engine as well as its Ethernet ports. A standard NIC moves packets between the cable and the server’s memory and leaves almost everything else to the host processor. A SmartNIC can inspect, change, encrypt or steer those packets itself, before the host ever sees them.
The point is to free the CPU. At 100 Gb/s and above, the work of running virtual switches, tunnels, firewalls and encryption can take a large share of a server’s cores. Moving that work onto the card returns those cores to the applications, and the card does it at line rate with lower and more predictable latency.
SmartNICs plug into a PCIe x8 or x16 slot, or an OCP NIC 3.0 slot in a rack server, and connect to the network through SFP28, QSFP56 or QSFP112 cages.
Types of SmartNIC
“SmartNIC” covers a wide range of cards. They are usually grouped by what does the processing.
-
//
ASIC-based. Fixed-function offload engines built into the network chip, such as the RDMA, overlay and crypto engines on NVIDIA ConnectX adapters. Fast and efficient, but limited to the functions the silicon was designed for.
-
//
FPGA-based. A field-programmable gate array sits in the data path and can be reprogrammed for custom packet processing, filtering or timestamping. Very flexible and low-latency, but harder to program.
-
//
SoC-based, or DPU. A system-on-chip with general-purpose Arm cores, memory and accelerators, running its own Linux. NVIDIA BlueField, AMD Pensando and Intel’s IPUs are examples. These are the most capable, and the most power-hungry.
SmartNIC vs NIC vs DPU
The boundaries are blurred because vendors use the terms differently, but the table shows the usual distinction.
| Feature | Standard NIC | SmartNIC | DPU |
|---|---|---|---|
| Processing on the card | Basic offloads: checksums, segmentation, receive-side scaling | Hardware engines or FPGA logic for specific offloads | Arm cores plus hardware accelerators |
| Programmable | No | Partly, through firmware or FPGA images | Yes, runs its own operating system |
| Typical offloads | None beyond the basics | RDMA, overlay tunnels, crypto, packet steering | Virtual switching, storage, firewall and security services |
| Typical speeds | 1 to 25 Gb/s | 25 to 400 Gb/s | 100 to 400 Gb/s |
| Typical power | A few watts | Around 15 to 75 W | Around 75 to 150 W |
| Typical host | Industrial PCs, desktops, most servers | Data centre and edge servers | Cloud and AI data centre servers |
A DPU, or data processing unit, is best read as the top tier of SmartNIC. The name was popularised by NVIDIA for its BlueField cards, and Intel uses infrastructure processing unit (IPU) for the same idea. What sets a DPU apart is that it can run the server’s infrastructure services, such as the virtual switch, storage stack and security policy, completely separately from the host operating system.
What SmartNICs are used for
Most SmartNIC work falls into a few groups.
-
//
Virtual networking. Running Open vSwitch and VXLAN or GENEVE overlay tunnels on the card, so virtual machines and containers get full network speed without loading the host.
-
//
RDMA and AI clusters. Moving data directly between server or GPU memory over RoCEv2 or InfiniBand. Our explainer on how RDMA speeds up edge AI covers the GPU side, and RoCEv2 explained shows how it runs over Ethernet.
-
//
Security. IPsec, TLS and MACsec encryption, stateful firewalls and denial-of-service filtering at line rate, isolated from the host.
-
//
Storage. NVMe over Fabrics, so remote flash appears as a local drive, with compression and encryption handled on the card.
-
//
Precise timing. Hardware timestamping and PTP for 5G radio access networks, financial trading and synchronized sensor capture.
SmartNICs at the edge
Most industrial PCs and embedded systems do not need a SmartNIC. Their standard Intel or Realtek controllers already include basic offloads, and 1 to 10 Gb/s of traffic is easily handled by the processor. SmartNICs belong in edge servers that aggregate many camera streams, run a private 5G core, or act as a secure gateway between an operational network and the cloud.
A simple test is to measure how much processor time goes on networking. If virtual switching, encryption or packet capture is using several cores at peak load, or latency varies too much for the application, a SmartNIC is worth considering. If the network load is light, a standard NIC and the cores you already have are the simpler choice.
They need the right host: a free PCIe x8 or x16 slot with enough airflow, as these cards run hot. GPU servers such as the Advantech SKY-622G4 set aside PCIe 5.0 x16 slots for NICs, SmartNICs or DPUs alongside the GPUs, and our rack mount servers and edge AI servers cover the range from short-depth edge systems to multi-GPU builds.
Quick answers on SmartNICs
SmartNICs: common questions
What is a SmartNIC used for?
Offloading network, security and storage work from the CPU. Typical jobs are virtual switching, encryption, RDMA and NVMe over Fabrics, done on the card at full line rate.
Is a SmartNIC the same as a DPU?
A DPU is a type of SmartNIC. Every DPU is a SmartNIC, but a DPU adds general-purpose Arm cores and runs its own operating system, so it can host whole infrastructure services.
What is the difference between a SmartNIC and a regular NIC?
Processing on the card. A regular NIC moves packets and does basic checksum work. A SmartNIC can process, encrypt and steer traffic itself.
Who makes SmartNICs?
Mainly NVIDIA, AMD, Intel and Broadcom. NVIDIA sells ConnectX and BlueField, AMD sells Pensando and FPGA-based Alveo cards, Intel sells IPUs and Broadcom sells Stingray adapters.
Does a SmartNIC need a special slot?
No, but it needs a fast one. Most use a standard PCIe Gen4 or Gen5 x8 or x16 slot, or an OCP NIC 3.0 slot, plus good airflow.
We fit and check network cards, including SmartNICs where a design calls for them, as part of our standard server build, matched to a slot with the lanes and cooling they need.