"" names one thing at the top and dozens at the bottom. Every Ethernet link carries the same frames, but the physical layer underneath (the PHY: the cable, the light or electrical signal, the way bits are coded onto it) has been reinvented for each new speed and medium. A 1 Gb/s link on copper and a 100 Gb/s link on fibre share almost nothing below the frame. Knowing which PHY is in play explains what cable a link needs, how far it can reach, and why it sometimes comes up at the wrong speed.

Reading the name

IEEE 802.3 names its physical layers in a compact code, and David Law, who chaired the working group, has written the key down (Law, IEEE 802.3, 2009):

  • the number is the speed: 10, 100, 1000, 10G, 100G, 400G;
  • BASE means baseband: one signal on the medium, not several modulated carriers;
  • the letters after the hyphen describe the medium and the coding: T for twisted pair; X for block coding taken from another standard (4B/5B, 8B/10B); R for "scrambled coding (64B66B)"; for fibre, S, L and E for short, long and extra-long wavelength (850, 1310 and 1550 nanometres);
  • a final number counts lanes: "where the data rate is 10Gb/s or more the number (m) represents the number lanes, the absence of a number indicating single lane operation".

A later list from the same author gives the fibre letters reach classes too, S for about 100 m, L for 10 km, E for 40 km, and adds D, "parallel single mode (500m)", and F, "fiber (2 km)" (Law, PHY names, 2015). So 100GBASE-SR4 reads as 100 Gb/s, baseband, short-reach multimode fibre, four lanes; and indeed Cisco's data sheet describes its 100GBASE-SR4 module as "100m over OM4 MMF" with an MPO connector (Cisco, 100G QSFP data sheet).

A short history in amendments

Each speed arrived as an amendment to IEEE 802.3, approved on a date the task force pages still record:

AmendmentApprovedWhat it brought
802.3i28 September 199010BASE-T, 10 Mb/s on twisted pair
802.3u1995100BASE-T, Fast Ethernet
802.3x20 March 1997Full duplex and pause flow control
802.3zJuly 1998Gigabit Ethernet on fibre (1000BASE-X)
802.3abJune 19991000BASE-T, Gigabit on twisted pair
802.3aeJune 200210 Gigabit Ethernet on fibre
802.3an8 June 200610GBASE-T
802.3baJune 201040 and 100 Gigabit Ethernet
802.3by30 June 201625 Gigabit Ethernet
802.3bs6 December 2017200 and 400 Gigabit Ethernet
802.3df16 February 2024800 Gigabit Ethernet

The dates are from each task force's page (802.3i, 802.3z, 802.3ab, 802.3ae, 802.3an, 802.3ba, 802.3by, 802.3bs, 802.3df, 802.3x); 802.3u's year is from the working group's own overview (Law, 2011).

Copper: 100 metres, and the category that gets you there

Every twisted-pair PHY from 10BASE-T to 10GBASE-T aims at the same 100-metre reach, the length of a horizontal run in a structured cabling plant; what changes is the cable it needs.

PHYCableReach
10BASE-T"Category 3 or Category 5 ""100 m (328 ft.)" (Cisco, Ethernet basics)
100BASE-TXtwo pairs of Category 5100 m (Law, 2011)
1000BASE-Tall four pairs of Category 5"100 meters of Category 5 balanced cabling" (802.3ab tutorial, 1998)
10GBASE-TCategory 6A (55 m on unshielded Category 6)"up to 100 meters" (Cisco, 10GBASE-T white paper)

For new installations, Fluke Networks' advice is plain: a new cabling system "should be at least category 5e ANSI/-C" (Fluke Networks). How a plant is laid out to keep every run inside its 100 metres is the subject of structured cabling, and the cable run planner does the arithmetic.

How bits become signals

A PHY's other job is coding: turning a stream of bits into a signal the medium can carry and the receiver can recover. The codes changed with every generation:

  • Manchester coding for 10BASE-T: "10Base-T uses Manchester encoding" (UNH-IOL), a transition in every bit, simple and wasteful.
  • 4B/5B and MLT-3 for 100BASE-TX: the encoder "converts 4-bit (4B) nibble data generated by the MAC into 5-bit (5B) code-groups", and a three-level MLT-3 signal carries them (Texas Instruments, DP83822 data sheet).
  • -5 on four pairs for 1000BASE-T: five signal levels, and "each pair a full duplex channel supporting effective data rate of 250Mbp/s in both directions simultaneously" (802.3ab tutorial). Gigabit on copper sends and receives on the same pairs at once.
  • 8B/10B and 64B/66B: the X and R of the names above. 64B/66B is the coding of the BASE-R family; it is not universal at 10 Gb/s, since 10GBASE-T, the copper PHY, codes differently.
  • PAM-4: as lanes passed 25 Gb/s, Ethernet "transitioned from non-return-to-zero (NRZ) used for 25Gb/s per lane to four level Pulse-amplitude modulation (PAM-4) for 50 Gb/s per lane and above" (Ethernet Alliance, 2022 roadmap): two bits per symbol instead of one.

Fibre, and speed in lanes

On fibre the names describe reach and wavelength. The working group's overview lists 1000BASE-SX at 220 or 550 metres on multimode fibre, 1000BASE-LX at 5 km, 10GBASE-SR at up to 300 m on multimode, 10GBASE-LR at 10 km and 10GBASE-ER at 40 km on (Law, 2011). Above 10 Gb/s, speed comes from lanes as much as from faster signals: 40GBASE-SR4 runs "4 lanes, 850 nm MMF" (Cisco, 40G QSFP); 400G modules use eight multimode pairs at up to 100 m (SR8, which the objective describes as "400 Gb/s operation over 8 pairs of MMF with lengths up to at least 100m") or four wavelengths on one single-mode pair to 2 km (FR4) (IEEE 802.3cm objectives; IEEE 802.3cu objectives; Cisco, 400G data sheet). The fibre itself, its grades and connectors and its loss budget, is the subject of fibre basics.

Full duplex, and the end of collisions

The original Ethernet was a shared medium: one station transmitted at a time, collisions were detected and resolved, and half duplex was the only mode. On that medium, Cisco reminds readers, "collisions are not bad; they are essential to correct Ethernet operation" (Cisco, collisions). Switching gave each link to two stations only, and 802.3x, approved in 1997, added "full duplex mode of operation on a speed-independent basis" with "pause-based flow control" (IEEE 802.3x). "In full-duplex Ethernet, collision detection is disabled" (Cisco, collisions). Gigabit Ethernet still defined half duplex on paper, but from 10 Gb/s on, every task force wrote the same objective: "Support full-duplex operation only" (IEEE 802.3ae objectives).

Auto-negotiation

Twisted-pair Ethernet settles speed and duplex automatically. Cisco describes it as "an optional function of the IEEE 802.3u Fast Ethernet standard that enables devices to automatically exchange information over a link about speed and duplex abilities" (Cisco, auto-negotiation). It is defined in Clause 28 and carried by Fast Link Pulses: "An FLP burst is simply a sequence of 10BASE-T Normal Link Pulses", sent every "16ms +/- 8ms" (UNH-IOL, Clause 28), and its base page also advertises pause ability for flow control (IEEE 802.3bz ad hoc, 2015).

From Gigabit on copper, it is no longer optional. The standard says "All 1000BASE-T PHYs shall provide support for (Clause 28)", and an IEEE interpretation confirms that "selecting 1000BASE-T mode of operation still requires that Auto-Negotiation be used", because negotiation is also where the two ends settle the MASTER-SLAVE relationship, "necessary for establishing the timing control of each PHY" (IEEE 802.3 interpretation 2-07/05). The task force behind 2.5GBASE-T and 5GBASE-T proposed the same rule for them (IEEE 802.3bz ad hoc, 2015), and auto-negotiation support was an objective of 10GBASE-T from the start (IEEE 802.3an objectives). Fibre is different: in 1000BASE-X "speed selection is not possible through Auto-negotiation" (interpretation, as above), and some 10 Gb/s fibre adapters "do not support Auto_Negotiation mode because they only work at one speed" (IBM).

Parallel detection, and the duplex mismatch

The classic fault comes from one side negotiating and the other not. A port set to auto that hears no negotiation from its neighbour falls back on parallel detection, "an auto-negotiating device's means to establish links with non-negotiating, fixed speed devices": it can tell the speed from the signal, but "a device can never parallel detect to a full duplex link" (UNH-IOL). So it assumes half duplex. Cisco's table of combinations shows the result: one end fixed at "100 Mbps, Full-duplex", the other at "AUTO", which settles at "100 Mbps, Half-duplex", and the comment "Duplex Mismatch" (Cisco, document 17053).

A mismatched link does not fail; it limps. "The late collision is seen on the half-duplex side. As the half-duplex side transmits, the full duplex side does not wait its turn and transmits simultaneously", and the counters fill with a "rapid increase of and alignment errors on the half-duplex side, and runts on the full-duplex port" (Cisco, auto-negotiation). The symptoms Cisco lists are "slow performance, intermittent connectivity, data link errors, and other issues" (document 17053), worst under load, which is why the fault can go unnoticed for months on a lightly used port.

The fix is a rule about both ends, never one: "Always set both sides of the connection to auto- speed/duplex or set the speed/duplex manually on both sides" (Cisco, document 12027), and Cisco's current advice is "to leave auto-negotiation on for those devices compliant with 802.3u" (Cisco, auto-negotiation). Hard-setting one end and leaving the other on auto is the one configuration that is guaranteed to go wrong.

What to remember

  • Read the name: speed, BASE, medium and coding letters, lane count. 100GBASE-SR4 is short-reach multimode, four lanes.
  • Twisted pair reaches 100 metres at every speed; the cable category is what changes, Category 6A for 10GBASE-T.
  • Full duplex ended collisions; from 10 Gb/s on, Ethernet is full duplex only.
  • Auto-negotiation is mandatory for Gigabit and faster on copper. Fibre at a fixed speed does not negotiate it.
  • Duplex mismatch: one end fixed, the other auto. Late collisions and FCS errors on the half-duplex side, runts on the full-duplex side. Configure both ends the same way.

Sources