XGMII, XAUI, XFI/SFI and the AUI family: 10G to 800G
From 10 Gigabit onward the interface between host and module is a bundle of serial lanes, and its name tells you the rate, the lane count and the modulation: SFI is one 10G NRZ lane, CAUI-4 four 25G lanes, 400GAUI-8 eight 50G PAM4 lanes, 800GAUI-8 eight 100G PAM4 lanes. The module must accept exactly that bundle, retime or amplify it, and map it onto its optical lanes — which may be a different number. This page lists the interfaces, the chip-to-chip vs chip-to-module distinction with its compliance points, the OIF CEI specifications behind them, the electrical-to-optical lane mapping, and who does the signal conditioning on each generation.
10 Gigabit
| Interface | Form | Where | Notes |
|---|
| XGMII (Cl. 46) | 32-bit data + 4-bit control each way, DDR at 156.25 MHz, 74 signals | on-chip between MAC and PCS | ≤ 7 cm; not on boards today |
| XAUI (Cl. 47) | 4 lanes × 3.125 GBd, 8b/10b | MAC/XGXS ↔ PHY; XENPAK, X2, XPAK modules | up to 50 cm PCB; obsolete for modules |
| RXAUI | 2 × 6.25 GBd | ASIC ↔ PHY (Marvell) | pin-saving XAUI |
| XFI (XFP MSA) | 1 × 10.3125 GBd, 64b/66b | host ↔ XFP | module contains CDR — a retimed interface (XFP) |
| SFI (SFF-8431) | 1 × 10.3125 GBd, 64b/66b | host ↔ SFP+ | module usually limiting or linear without CDR; host SerDes does equalisation; compliance points B (host output), C (module output), D at the host input (SFP+) |
| 10GBASE-KR (Cl. 72) | 1 × 10.3125 GBd with link training | backplane; ASIC ↔ 10GBASE-T PHY | AN Clause 73 |
| USXGMII | 1 × 10.3125 GBd, rate-adapted | ASIC ↔ multigig PHY | SGMII & serial gigabit |
SFI's "linear" option carried 10GBASE-LRM: the module passes the distorted multi-mode signal linearly and the host's electronic dispersion compensation (EDC) cleans it — the ancestor of today's LPO idea.
25, 40, 50 Gigabit
| Interface | Lanes | Standard | Cage |
|---|
| 25GAUI | 1 × 25.78125 GBd NRZ | Annex 109A (C2C), 109B (C2M) | SFP28 |
| XLAUI | 4 × 10.3125 GBd | Annex 83A/B | QSFP+ (40GBASE-LR4 etc.) |
| XLPPI | 4 × 10.3125, non-retimed parallel physical interface | Annex 86A | QSFP+ SR4/CR4 without CDR |
| 50GAUI-2 | 2 × 26.5625 GBd NRZ | Annex 135B/C | QSFP28 (2 × 50G breakout), SFP56 (some) |
| 50GAUI-1 | 1 × 26.5625 GBd PAM4 | Annex 135D/E | SFP56, 50GBASE-SR/FR/LR modules |
100 Gigabit
| Interface | Lanes | Standard | Cage |
|---|
| CAUI-10 | 10 × 10.3125 GBd | Annex 83A/B | CFP, CXP, 100GBASE-SR10 |
| CAUI-4 | 4 × 25.78125 GBd NRZ | Annex 83D (C2C), 83E (C2M) | QSFP28, CFP2/CFP4 |
| CPPI-4 | 4 × 25.78, non-retimed | Annex 86A-style for SR4 | QSFP28 SR4 without CDR (rare) |
| 100GAUI-4 | 4 × 26.5625 GBd NRZ (with RS-FEC(544) framing) | Annex 135D/E | QSFP28 in 100GBASE-CR4/KR4 KP4 mode |
| 100GAUI-2 | 2 × 53.125 GBd PAM4 | Annex 135F/135G (C2M) | QSFP56 100G lanes, 2 × 100G breakout, DR/FR1 modules with 2-lane hosts |
| 100GAUI-1 | 1 × 106.25 GBd PAM4 | Annex 120F/120G (C2M) (802.3ck) | QSFP112, OSFP 100G-per-lane, SFP112 |
200, 400, 800 Gigabit
| Interface | Lanes | Standard | Cage |
|---|
| 200GAUI-8 | 8 × 26.5625 GBd PAM4 | Annex 120B/C | QSFP-DD in 200G mode (rare) |
| 200GAUI-4 | 4 × 53.125 GBd PAM4 | Annex 120D/120E (C2M) | QSFP56, OSFP/QSFP-DD 2 × 200G |
| 200GAUI-2 | 2 × 106.25 GBd PAM4 | Annex 120F/G | QSFP112 2 × 200G, OSFP |
| 400GAUI-16 (CDAUI-16) | 16 × 26.5625 GBd NRZ | Annex 120B/C | CFP8 |
| 400GAUI-8 (CDAUI-8) | 8 × 53.125 GBd PAM4 | Annex 120D/120E (C2M) | QSFP-DD, OSFP (400G) |
| 400GAUI-4 | 4 × 106.25 GBd PAM4 | Annex 120F/120G (C2M) | QSFP112, OSFP 400G, half of an 800G port |
| 800GAUI-8 | 8 × 106.25 GBd PAM4 | 802.3df annexes (C2C/C2M) | OSFP, QSFP-DD800 |
| 800GAUI-4 / 1.6TAUI-8 | 4 / 8 × 212.5 GBd PAM4 | 802.3dj (in progress) | OSFP224, OSFP-XD, QSFP-DD1600 |
C2C vs C2M and compliance points
| Chip-to-chip (C2C) | Chip-to-module (C2M) |
|---|
| Path | ASIC ↔ ASIC/PHY on the same board or across a connector | ASIC → host PCB → cage connector → module PCB → module IC |
| Reach | up to ~25–50 cm (NRZ), less at PAM4 | ~10–15 cm host trace + connector |
| Test points | TP0/TP5 at chip pins | TP1 (host output at module connector), TP1a (after a reference channel = module input), TP4 (module output at connector), TP4a (at host input after host channel) |
| Module role | — | must meet output spec at TP4 (eye, jitter, TDECQ-like metrics) and tolerate host signal at TP1a with the specified stressed eye |
| Retiming | both chips retime | module retimes (CDR/DSP) — or is linear (LPO) and relies on the host |
Host boards are qualified against the C2M reference channel; a long or lossy host trace (high-radix 800G switches) is why retimers and gearboxes appear on switch boards (Inside a switch).
OIF CEI behind the IEEE annexes
| OIF CEI | Rate / modulation | IEEE equivalent | Reach class |
|---|
| CEI-28G-VSR | 28 GBd NRZ | CAUI-4 C2M | very short reach (module) |
| CEI-28G-SR / -MR / -LR | 28 GBd NRZ | CAUI-4 C2C / backplane | short / medium / long (backplane) |
| CEI-56G-VSR-PAM4 | 56 GBd PAM4 | 400GAUI-8 / 200GAUI-4 C2M | module |
| CEI-56G-MR/LR-PAM4 | 56 GBd PAM4 | 200GBASE-KR4 class | backplane |
| CEI-112G-VSR-PAM4 | 112 GBd PAM4 | 400GAUI-4 / 800GAUI-8 C2M | module |
| CEI-112G-XSR / -MR / -LR | 112 GBd PAM4 | die-to-die / C2C / backplane | — |
| CEI-112G-Linear-PAM4 | 112 GBd PAM4 | LPO host interface | module, linear |
| CEI-224G (in progress) | 224 GBd PAM4 | 800GAUI-4 / 1.6TAUI-8 | — |
OIF publishes the electrical envelopes; IEEE annexes reference or align with them.
Electrical lanes vs optical lanes
| Module | Electrical (host) | Optical (line) | Conversion in module |
|---|
| 10GBASE-LR SFP+ | 1 × 10G | 1 λ × 10G | none |
| 40GBASE-LR4 | 4 × 10G | 4 λ × 10G | none (WDM mux) |
| 100GBASE-SR4 / LR4 | 4 × 25G NRZ | 4 fibres / 4 λ × 25G | none |
| 100GBASE-DR / FR1 / LR1 | 4 × 25G NRZ | 1 λ × 100G PAM4 | gearbox + PAM4 DSP (4:1) |
| 100G on QSFP112 (100GAUI-1) | 1 × 100G PAM4 | 1 × 100G PAM4 | retime only |
| 200GBASE-FR4 | 4 × 50G PAM4 | 4 λ × 50G PAM4 | retime |
| 400GBASE-DR4 / FR4 | 8 × 50G PAM4 | 4 fibres / 4 λ × 100G PAM4 | gearbox 2:1 in DSP |
| 400GBASE-SR8 | 8 × 50G PAM4 | 8 fibres × 50G PAM4 | retime |
| 400G on 400GAUI-4 hosts | 4 × 100G PAM4 | 4 × 100G PAM4 | retime (or LPO linear) |
| 800G-DR8 / 2×FR4 | 8 × 100G PAM4 | 8 fibres / 2 × 4 λ × 100G | retime |
| 800G-DR4 (200G/λ) | 8 × 100G PAM4 | 4 × 200G PAM4 | gearbox 2:1 |
| 1.6T-DR8 (OSFP224) | 8 × 200G PAM4 | 8 × 200G PAM4 | retime |
The gearbox is why a 400G-DR4 module draws 8–10 W and a 400G-SR8 less, and why CMIS lists separate host and media interface IDs per application (CMIS).
Who conditions the signal
| Generation | In the module | On the host |
|---|
| 10G SFP+ (SFI) | limiting amp or linear TIA; no CDR on most | SerDes with CTLE/DFE equalisation; EDC for LRM |
| 10G XFP (XFI) | CDR | plain SerDes |
| 25G SFP28, 100G QSFP28 | CDR on Tx and Rx (bypassable: SFF-8636 bytes 98/99) | SerDes FFE/DFE; pre-emphasis programmable in the module's control bytes (SFF-8636 page 03h 234–241: Rx output amplitude/emphasis, Tx input equalisation) |
| PAM4 200G–800G | DSP retimer with FFE/DFE/MLSE; host-side equaliser settings via CMIS (page 10h/11h Tx input EQ, Rx output amplitude/pre-cursor/post-cursor) | SerDes DSP |
| LPO 400G/800G | linear driver and TIA, no DSP | the switch ASIC's SerDes DSP compensates the whole channel; host must be LPO-capable and directly attached (DSP & LPO) |
| Retimed vs linear AEC/ACC cables | AEC: DSP in the plug; ACC: linear amplifier | as above (Cable internals) |
Rate selection and CDR bypass controls: Rate & application selection, CDR.
In CodingBox
For SFF-8636 modules CodingBox shows the CDR-present bits, rate-select capability and the equalisation/emphasis control bytes; for CMIS modules the application table with host interface IDs (CAUI-4, 400GAUI-8, 100GAUI-1…) against media interface IDs — the declaration a host uses to pick lane count and modulation. A module advertising only 400GAUI-8 will not run on a 400GAUI-4 (QSFP112-class) port, however good its optics (Check transceiver, CMIS upper pages).