Ethernet over copper (twisted pair) §

A twisted-pair link is one jacket containing four twisted pairs of 0.4–0.6 mm copper, terminated in an 8P8C modular jack (“RJ45”): 8 pins, each pair on a neighbouring pin pair, with one pair used for transmit and one for receive in the 10/100BASE-T pinout, and all four used bidirectionally in 1000BASE-T. Because each pair is transformer-coupled at both ends, the same four pairs can simultaneously carry Power over Ethernet (PoE): the power source injects a common-mode voltage between the pairs (up to 90 W with IEEE 802.3bt), which the data transformers do not see, while the data rides the differential signal on each pair. This is copper’s quiet killer feature: one cable delivers both the bits and the watts to cameras, access points, and phones — something no fibre can do.

The category rating is analog: how far the cable attenuates a sine wave of a given frequency, and how much a pair’s signal leaks into its neighbours (crosstalk) and reflects back (return loss). Turning that into a data rate is the PHY’s job — which is why the same Cat 6a cable carries 1, 2.5, 5, or 10 Gb/s depending on which transceivers you plug in. The PHY spends its noise budget on:

• Cancellation: 10GBASE-T actively subtracts echo, near-end crosstalk (NEXT), and far-end crosstalk (FEXT) from all 7 other wire-combinations in the cable (see the next section).
• Coding: multi-level symbols (PAM16, see below) pack more bits per Hz.
• Forward error correction: LDPC codes repair the errors that survive.

Better cable (higher category) buys margin for all three — which is why the same PHY reaches the full 100 m on Cat 6a but only its certified 55 m limit on Cat 6 (table above).

What the extra rating buys physically is geometry: tighter, deliberately unequal twist pitches per pair (so that crosstalk between two pairs does not accumulate in phase over the whole run), a cross-shaped plastic spline separator that keeps the four pairs apart in Cat 6a, and — in the shielded categories — foil around each pair and a braid or foil around the whole bundle, which converts crosstalk from a computation problem into a grounding problem.

10GBASE-T must fit 2.5 Gb/s into each pair’s usable bandwidth (~400 MHz Nyquist for 800 MBd). That is 6.25 bits per symbol-period — impossible with 2-level signalling (1 bit) or even PAM4 (2 bits). The trick is to treat two consecutive PAM16 symbol intervals on a pair as one 2-D symbol: the 16×16 grid of possible (level, next level) pairs has 256 points, and the DSQ128 constellation keeps only the 128 checkerboard ones, which maximises the minimum distance between points and roughly halves the power needed for the same error rate. (So the “two dimensions” are time — two successive levels on the same wire — not two wires or two carriers.)

Tomlinson–Harashima precoding moves part of the equalisation burden to the transmitter: the sender pre-distorts the signal using its knowledge of the channel (learned during link training), so the receiver’s equaliser can be simpler and the constellation points land where expected. This matters because a 10GBASE-T receiver is already doing echo, NEXT, and FEXT cancellation on every sample (next section).

Example. Compare the two generations on the same four-pair cable: 1000BASE-T needs 2 bits per pair per symbol period and gets them from 5 levels in one dimension; 10GBASE-T needs 3.125 information bits per PAM16 symbol and gets them from a 2-D constellation of 128 points spanning two symbol periods. Each doubling of bits-per-symbol costs roughly 3 dB of signal-to-noise ratio — which is why 10GBASE-T spends its entire noise budget on cancellation, precoding, and FEC before it can afford the constellation at all.

All three interferers are deterministic given the channel: the cable is a linear, time-invariant system over the life of a link, so each interferer is a filtered copy of a signal the PHY chip itself knows (it generated the transmit symbols on all four pairs, and it measured the channel during link training). The receiver therefore runs one adaptive filter per interferer — 7 per wire, 28 per link — subtracting each filtered copy from the received sample before the slicer decides which constellation point was sent.

• Echo is the strongest term (the culprit is metres away, at the connector) and needs the longest filter; it is the same problem a voice modem solves with its hybrid, but at 800 MBd.
• NEXT is strong too (the interfering transmitter sits in the same chip, metres away) and is continuous, so the canceller runs on every sample.
• FEXT is the weakest term, because the interfering signal travelled the whole cable before leaking in — which is why FEXT cancellation is the first to be relaxed in the cheaper 2.5/5GBASE-T PHYs.

What cannot be cancelled this way is alien crosstalk: interference from transmitters in other cables bundled in the same tray, whose symbols the PHY chip has no knowledge of. Alien crosstalk is the reason the certified reach limits in the table above are shorter than a within-the-cable analysis suggests, and the reason Cat 6a’s specification added alien-crosstalk limits (TIA TSB-155 documented the interim Cat 6 limits).

Autonegotiation is often confused with link training, but they answer different questions:

• Negotiation (the FLP exchange) is a policy step, done at the physical layer with analogue pulses precisely so it works when nothing else about the link is configured: it answers “what rates do we both support?” It uses a priority table (10G over 5G over 2.5G over 1G, full over half duplex) so the outcome is deterministic even if both sides advertise overlapping sets. A port that supports 1/2.5/5/10 GbE talking to a 1 Gb/s-only printer settles on 1 Gb/s without any configuration.
• Training is a physics step, done after the rate is chosen: the two PHYs send known sequences so each receiver can learn the actual channel — its echo, NEXT, and FEXT impulse responses — and converge the cancellers. Training is where the real cable is measured, and it is the step that can fail: if the measured margin at the negotiated rate is too thin (a long run on old cable, a bundle full of alien crosstalk), the PHYs fall back and renegotiate one rate down. This is why a 10GBASE-T port on a marginal Cat 6 run does not sit there half-broken — it comes up at 2.5 or 5 Gb/s.

The practical consequence: the cable category, not the port speed, sets the ceiling. The negotiation advertises what the silicon can do; training measures what the copper can do; the link runs at the lower of the two.

Copper wins on cost, power, and convenience (power over Ethernet, existing building wiring); fibre wins on reach, immunity to electrical noise, and bandwidth ceiling. The crossover is around 100 m — which is exactly why 10GBASE-T stops there. The physics is unforgiving on both sides: copper’s attenuation grows roughly as $\sqrt{f}$ per metre (skin effect — see Cabled external interconnects), while fibre’s attenuation is a flat ~0.2–3 dB/km with no frequency-dependent wall below hundreds of GHz. So every doubling of a copper PHY’s rate must be paid for in distance, in cable quality, or in silicon (DSP, FEC) — fibre pays none of that tax until reach is measured in kilometres.

Per-PHY fibre reaches vary widely — 10GBASE-SR alone spans roughly 26–400 m depending on the fibre grade (OM3 vs OM4), short-reach copper twinax (10GBASE-CR) sits below even that at ~7 m, and 10GBASE-ER reaches 40 km — but the order of magnitude is what matters: fibre begins where copper’s 100 m wall ends.

See Optical interconnects for the fibre side, and Board-level interconnects for the chip-to-chip link (XFI) between the MAC and the PHY itself.

A switch port and a NIC both support 1/2.5/5/10 GbE (NBASE-T). They are connected by 60 m of Cat 6 cable. Which rate will they settle on, and why?