NIC ports, speeds and PCIe bandwidth
Đội ngũ của chúng tôi đang thực hiện bản dịch. Bài viết này tạm thời chưa có bản tiếng Việt và đang được hiển thị bằng tiếng Anh.
A NIC's front panel promises "2 × 100G"; whether the server can actually move 200 Gb/s depends on the PCIe slot behind the card, and whether a given module works depends on the cages and the controller's lane map — exactly as on a switch, but with fewer ports and a bus that is easy to under-provision. This page lists the cages and speeds found on NICs, the rules for dual-rate ports and breakout cables, the PCIe arithmetic, the OCP 3.0 form factor, and the power and cooling limits that matter for optics on cards.
Cages and speeds on NICs
| Cage | Speeds seen on NICs | Typical controllers | Notes |
|---|---|---|---|
| SFP+ | 10G (1G on some) | Intel X520/X710, Broadcom 5741x | 1G support varies; check the card's spec |
| SFP28 | 25G, 10G (1G rare) | Intel XXV710/E810-XXV, ConnectX-4 Lx/-6 Lx, Broadcom P225 | 25G needs FEC agreement (FEC & AN) |
| SFP56 | 50G, 25G, 10G | ConnectX-6 Lx/Dx (some) | PAM4 lanes |
| QSFP28 | 100G, 4 × 25G, 2 × 50G, 40G, 4 × 10G | Intel E810-CQDA2, ConnectX-5/6, Broadcom P2100 | breakout via cable only where the controller allows |
| QSFP56 | 200G, 100G, 2 × 100G | ConnectX-6 Dx/-7, Broadcom P2200, Pensando | — |
| QSFP112 | 400G, 200G, 100G, 4 × 100G | ConnectX-7, Broadcom P1400GD | 100G/lane; accepts QSFP28/56 |
| QSFP-DD | 400G, 2 × 200G, 4 × 100G | Broadcom BCM57608 | — |
| OSFP (twin-port) | 2 × 400G, 800G | ConnectX-7/-8, BlueField-3 | one cage, two logical ports on Nvidia twin-port designs |
| RJ45 | 1G / 10GBASE-T / mGig | X550, X710-T, BCM57416 | no module, PHY on the card |
Cage compatibility rules are the switch ones (Port types & cages).
Dual-rate ports and speed groups
- Most 25G NICs run each port independently at 10G or 25G — no quad grouping — but a few dual-port controllers share a SerDes group and force the same speed on both ports.
- 100G cards often support 40G modules; 200G/400G cards support 100G modules; going down more than one generation (1G in an SFP28, 10G in a QSFP56 via QSA) is rarely supported.
- Speed is usually auto from the module's compliance codes; forcing (
ethtool -s … speed) is needed when the module is dual-rate or unrecognised (Tools).
Breakout on NICs
| Case | Works? |
|---|---|
| 100G QSFP28 NIC port → 4 × 25G to a switch (DAC splitter) | only if the controller supports "4 × 25G port split" mode (many Nvidia and Broadcom cards do, most Intel do not); the card then shows 4 netdevs |
| Switch 100G port → 4 × 25G to four NIC SFP28 ports | yes — this is the common direction; the split is configured on the switch (Breakout & MPO) |
| 400G OSFP twin-port → 2 × 400G or 4 × 200G / 8 × 100G | Nvidia twin-port cards: 2 logical ports per cage; further splitting depends on firmware |
| 2 × 50G from a QSFP28 port | some controllers |
PCIe: the real port speed
| PCIe generation | Per lane (raw) | x8 usable (≈) | x16 usable (≈) |
|---|---|---|---|
| Gen 3 | 8 GT/s | 56–63 Gb/s | 112–126 Gb/s |
| Gen 4 | 16 GT/s | 112–126 Gb/s | 225–252 Gb/s |
| Gen 5 | 32 GT/s | 225–252 Gb/s | 450–504 Gb/s |
| Gen 6 | 64 GT/s | ~450–500 Gb/s | ~900–1 000 Gb/s |
Usable throughput is 85–90 % of raw after encoding and packet overhead. Match ports to lanes:
| Card | Ports | Needs |
|---|---|---|
| 2 × 25G | 50 Gb/s | Gen 3 x8 |
| 2 × 100G | 200 Gb/s | Gen 4 x16 (Gen 3 x16 saturates at ~126) |
| 1 × 200G / 2 × 100G Gen 4 | 200 Gb/s | Gen 4 x16 |
| 2 × 200G | 400 Gb/s | Gen 5 x16 |
| 1 × 400G | 400 Gb/s | Gen 5 x16 (or two x16 via "socket direct"/multi-host) |
| 2 × 400G | 800 Gb/s | Gen 6 x16 or two Gen 5 x16 |
A 2 × 100G card in a Gen 3 x8 slot links at 100G on both ports and delivers ~60 Gb/s total — the optics are fine, the bus is not. Check lspci -vv for the negotiated link width and speed (LnkSta), and BIOS bifurcation settings.
OCP NIC 3.0
| Item | SFF (small form factor) | LFF (large) |
|---|---|---|
| Size | 76 × 115 mm | 139 × 115 mm |
| PCIe | up to x16 | up to x32 (two x16) |
| Power | up to 80 W | up to 150 W |
| Ports | 2–4 × SFP28/QSFP28/QSFP56/QSFP112 | 2 × QSFP-DD/OSFP or more |
| Optics access | front of the module, tool-less | same |
OCP cards sit at the rear of the server with their own airflow channel; module cooling is usually better than on a PCIe card buried between GPUs.
Power and cooling for optics on cards
| Item | Limit / rule |
|---|---|
| PCIe slot power | 25 W (x1…x8 low profile), 75 W (x16); 400G cards use an 8-pin auxiliary connector |
| Per-cage budget on NICs | SFP28: 1.5 W (level II); QSFP28: 3.5–4.5 W; QSFP56/112: 5–8 W; QSFP-DD/OSFP: 12–20 W — the card's datasheet states it (Power & consumption) |
| Airflow | 200–300 LFM across the card; optics on a card in a 1U server near the CPU exhaust run hot — DDM temperature above 60 °C is common |
| Thermal throttling | some controllers lower the module to low power or shut the port when the module reports high temperature |
| Coherent / ZR on NICs | not supported on standard NICs — class 8 exceeds the cage budget |
Multi-host and socket direct
High-end cards can present themselves to two CPU sockets (two x8 or x16 links) or to several hosts. The optical side is unchanged; the split is in PCIe. Relevant for 400G on Gen 4 platforms and for GPU servers where each GPU's NIC must sit on its own PCIe root.
Checklist before buying a NIC for optics
- Cage type and the speeds the controller supports at that cage.
- PCIe generation and lanes in the target slot vs total port bandwidth.
- Breakout needed? Controller must support port split.
- Module policy of the driver (Optics compatibility on NICs).
- Cage power budget vs planned modules; auxiliary power connector present.
- Airflow in the chassis position; consider OCP 3.0 for better cooling.
- OS/driver support for FEC modes and speed forcing.
In CodingBox
The module's compliance codes and power class tell you whether it can serve the NIC's cage and speed; CodingBox reads them before the module is installed, and the port-split and PCIe questions are settled on the server side (Check transceiver).