Network interface controller (NIC / Ethernet adapter card)

01

Set port speed, count and media

NVIDIA describes its ConnectX NICs as serving low-latency, high-throughput applications with one, two or four ports at 10, 25, 40, 50, 100, 200 and up to 400 Gb/s Ethernet: ConnectX-7 offers up to four ports and 400 Gb/s total, ConnectX-6 Dx up to two ports of 25, 50 or 100 Gb/s or one of 200 Gb/s using 50 Gb/s PAM4 SerDes, and ConnectX-6 Lx up to two 25 GbE ports or one 50 GbE port. [1]

At gigabit speeds on twisted-pair copper, a NIC uses an Ethernet PHY such as TI's DP83867, which supports 10BASE-Te, 100BASE-TX and 1000BASE-T to IEEE 802.3, up to 130 m of cable and IEEE 1588 start-of-frame detection, connecting to the MAC through MII, GMII or RGMII. [2]

Editorial calculation: a host slot must carry the ports' traffic — two 100 Gb/s ports in both directions need about 200 Gb/s each way, which a PCIe 4.0 x16 link (about 16 GT/s × 16 lanes, roughly 250 Gb/s raw per direction before encoding and protocol overhead) can carry, while an x8 link at half that cannot run both ports at full rate. [1]

The port speeds a NIC offers are defined in IEEE 802.3, the Ethernet standard. IEEE 802.3-2022 (approved 13 May 2022, published 29 July 2022, available free through the IEEE GET program) specifies Ethernet operation at selected speeds from 1 Mb/s to 400 Gb/s with one media access control (MAC) specification and management information base for all speeds; speed-specific media-independent interfaces let selected PHYs run over coaxial, twisted-pair or fibre cable or electrical backplanes, and it also covers power over selected twisted-pair PHY types. [3]

Editorial note: name each port's PHY type from IEEE 802.3 (for example 1000BASE-T copper or a specific optical or backplane type) as well as its speed, because cards of the same speed can use different media and cables. 800 Gb/s ports, quoted for some AI adapters, are later than the 400 Gb/s ceiling of the 2022 edition, so ask which amendment they follow. [3] [1]

Give the supplier the port speeds and count, media and connectors (RJ45 copper, SFP/QSFP optical or DAC), host interface and lane count, form factor and bracket, required offloads (RDMA/RoCE, encryption, storage, virtualisation), time-sync needs, operating system and driver support, power and cooling, temperature range and quantity.

02

Choose host interface, form factor, offloads and timing

ConnectX-6 Dx uses PCI Express 4.0 host connectivity; the ConnectX-6 Lx portfolio covers low-profile PCIe and Open Compute Project (OCP) 3.0-compliant cards across a range of form factors and speeds. NVIDIA lists hardware offloads including ASAP2 accelerated switch and packet processing, RDMA over Converged Ethernet (RoCE), GPUDirect Storage, NVMe over TCP acceleration (added in ConnectX-7), in-line TLS, IPsec and MACsec encryption and decryption at up to 400 Gb/s, DPDK performance and secure boot with a hardware root of trust. [1]

Starting with ConnectX-6 Dx, ConnectX NICs provide an accurate time-synchronisation service for data-centre applications; NVIDIA's Ethernet SuperNICs for AI workloads go up to 800 Gb/s. [1]

Editorial note: offloads only help if the operating system, drivers and applications use them, so list the software stack alongside the hardware features you require. [1]

03

Ask for receive queues, RSS and steering filters

The Linux kernel's networking documentation explains how a multi-queue NIC spreads received packets across CPUs with Receive Side Scaling (RSS): the NIC computes a hash over network and transport headers, for example a 4-tuple of IP addresses and TCP ports, and uses its low-order bits to index an indirection table whose entries hold queue numbers. The document says NICs should provide an indirection table at least 4 times larger than the queue count, and that a 4x table gives about 16% imbalance between the queues. Each receive queue has its own interrupt, signalled over PCIe with MSI-X, which can route each interrupt to a particular CPU. [4]

The document advises using no more receive queues than needed to keep any queue from overflowing a saturated CPU, because with interrupt coalescing the total interrupt work grows with each added queue. [4] Editorial note: queue counts, table sizes and filter features are set by the controller and its driver, so request them for the exact card and driver version alongside the offloads NVIDIA lists, such as RoCE and in-line encryption. [1]

04

Compare what suppliers offer

The companies in the manufacturer list below show this product on their own websites. Confirm the legal seller, factory and scope for your order.

Ask each supplier for the card's product brief or datasheet with port configurations, host interface, offloads, power, cooling airflow, operating temperature, dimensions and supported drivers and firmware lifecycle. [1]

05

Trace what each part controls

This is a functional reference built from the sources cited; it is not one maker's specification.

06

Fill the specification fields with values for your system

A NIC quotation needs values for your system. The table explains each field, gives published values where a reviewed source states them, and says what to obtain from the supplier.

07

Send the port, host and software data together

Paste it into your enquiry below or into an email to a maker, then replace every bracket.

Ports: [count; Gb/s]
Media: [RJ45; SFP; QSFP; DAC]
Host: [PCIe gen; lanes; OCP 3.0]
Offloads: [RoCE; crypto; NVMe-oF]
Timing: [PTP]
Software: [OS; drivers]
Power/temp: [W; °C; airflow]
Quantity: [units]

See manufacturer records Send an enquiry through CNFX