112G vs 224G Ethernet Signal Integrity: 802.3ck and 802.3dj

At a glance

What is 224G Ethernet?

Ethernet built on electrical lanes of 200 Gb/s each. The IEEE 802.3dj project defines 200 Gb/s, 400 Gb/s, 800 Gb/s and 1.6 Tb/s Ethernet over 1, 2, 4 and 8 such lanes (C-112), so a 1.6T port is eight of them. Its predecessor, 802.3ck, did the same with 100 Gb/s lanes for 100, 200 and 400 Gb/s (C-105). “112G” and “224G” are the industry’s names for these two lane classes, used for example by OIF’s CEI-112G and CEI-224G work (C-122, C-124). On the wire the lanes carry 106.25 and 212.5 Gb/s, data plus coding overhead.

Is 802.3dj approved?

Not yet. 802.3ck was approved on 21 September 2022 (C-105). 802.3dj was a draft in September 2026, at D3.2, going through the IEEE Standards Association ballot (C-111), and by mid-October 2026 OIF had not published a CEI-224G agreement either (C-124). This page therefore quotes the 802.3dj objectives and the baselines its task force adopted by vote, and marks them as draft: a later draft can still change a number.

112G vs 224G: what changed?

The symbol rate doubled and the modulation did not: both are PAM4 (C-106, C-115). The loss budget moved from a recommended 28 dB measured between the package balls to a 40 dB objective measured die to die (C-107, C-112). The error correction stayed the same Reed–Solomon code (C-110, C-115), and the error ratio it starts from was relaxed from 1e-4 to 2e-4 (C-116). The copper cable objective fell from at least 2 m to at least 1 m (C-109, C-113).

Reading the panel: the dots on the loss curve are the loss per inch at each rate’s Nyquist frequency, and the dashed line is the part the dielectric contributes. The dots on the pulses sit one symbol apart, 18.8 ps at 112G and 9.4 ps at 224G; every dot after the peak is a neighbouring symbol the pulse spills into. Through the same length the 224G pulse is lower and spills into more of its neighbours, which the receiver’s equaliser has to undo. The pulse shapes come from the site’s shared channel model, with each rate’s loss at its own Nyquist frequency and one fixed split between copper and dielectric loss, so the two pulses differ only in that loss and in the symbol time.

Twice the rate is not twice the loss

Loss in a board trace has two parts that grow differently with frequency. The dielectric turns a fixed fraction of the field’s energy into heat on every cycle, and an inch of trace holds twice as many cycles at twice the frequency, so its loss in dB grows in step with frequency: double the frequency, double the loss. Copper loss comes from the skin effect, which crowds the current into a layer whose depth shrinks as one over the square root of frequency, so the resistance, and the loss, grow only as the square root of frequency, about 1.41 times per doubling. Rough copper grows a little faster than that, but still slower than the dielectric part.

So moving from 112G to 224G multiplies the loss per inch by something between 1.41 and 2, depending on the mix, for a laminate whose Df, the dielectric’s loss tangent, is flat and copper that is smooth. Real low-loss laminates sit in the upper part of that range, because their Df also rises slowly with frequency: on the laminates in the panel the factor is about 1.6 to 2. A fixed budget then reaches about half to 70% as far, in practice a little over half. The loss budget calculator splits the two parts for any one frequency.

Absolute numbers depend on the board. One host board measured for a contribution to the task force, with a 6 mil trace in a 90 Ω pair, lost about 1.25 dB/in at 53.125 GHz at room temperature and about 1.40 dB/in hot, at 80 °C by its data files’ names (C-119). Its laminate is not named, and it is lower than any setting in the panel, whose laminates and first-order copper estimate are not that board. Use the panel for the ratio between the rates, and your own measured stackup for the dB.

Where the 40 dB is measured

A loss budget means little until you know its end points. 802.3ck recommends a backplane channel of at most 28 dB at 26.56 GHz from TP0 to TP5, test points at the package-to-board interface: the balls under each package (C-107). The packages sit outside that number and enter only the compliance calculation, COM (channel operating margin), as modelled package traces. OIF’s CEI-112G-LR, which shares 802.3ck’s method, measures its 28 dB test channel between the same package balls (C-123). 802.3dj’s objective is instead 40 dB at 53.125 GHz die to die, packages included, for both backplanes and copper cable (C-112, C-113). Only the copper cable link has a published split, so the like-for-like comparison below is for copper cable.

A like-for-like comparison tells more than “the budget grew by 12 dB”. 802.3dj’s copper cable budget splits as 11.5 dB for each host, package plus board, and 17 dB for the cable assembly (C-117). In a contribution’s nominal example each package takes 5 dB; its high-loss host takes 9 dB and is paired only with a nominal or low-loss host (C-118). Take two nominal packages out and 30 dB remains between the balls, or 26 dB when one end is the high-loss host, against 28.5 dB for a 112G copper cable link (C-109). In dB, the part on the board and in the cable stayed roughly where it was, now at twice the frequency, which fits the cable objective falling from 2 m to 1 m; the packages take 10 to 14 dB, a quarter to a third of the 40.

The 40 dB itself was a negotiation. The proposal for it reported that SerDes designers had asked for 36 to 37 dB and system builders for 42 to 45 dB, and called 40 the compromise (C-114). Whatever loss the channel is allowed, the receiver’s equaliser and the error-correcting code have to recover.

The long run moves into cable

If the board cannot carry 53 GHz far, carry it in something else. Worked examples contributed to the task force show how (C-118). A nominal host spends 6 dB on about 4.28 in of board trace, 1.4 dB per inch. The same host built with cable spends 2.8 dB on 2 in of board, 1 dB on a connector beside the package and 2.2 dB on 6.8 in of 30 AWG twinax, 0.32 dB per inch; a high-loss host runs 8 in of 32 AWG twinax for 3.2 dB, 0.40 dB per inch. Per inch at 53 GHz, the cable loses 3.5 to 4.3 times less than the board trace. That is the case for routing the long stretch of a lane over the board in cable, from a connector next to the package, rather than through it. The connectors and cables page covers the parts themselves.

The same code, a looser start

Both generations protect the data with RS(544,514), a Reed–Solomon code over 10-bit symbols (C-110, C-115). Its 30 parity symbols in every 544 let the receiver correct up to 15 wrong symbols per codeword. 802.3ck sized its links for a detector error ratio, the rate of wrong symbol decisions before correction, of 1e-4 (C-108); 802.3dj adopted 2e-4 for its backplane and copper cable links, twice as many (C-116). Keeping the code and relaxing the starting error ratio is another way the budget shifts work from the channel to the receiver. The NRZ vs PAM4 page explains why PAM4 needs the correction at all.

Side by side

OIF’s reach ladder at 112G

The OIF, an industry forum, writes implementation agreements for the electrical links between chips, modules and backplanes. Its CEI-112G family defines a ladder of reaches, all PAM4 at 36 to 58 GBd (C-122).

At 224G the OIF agreements were not yet published, and the 802.3dj backplane baseline planned to leverage OIF’s long-reach draft (C-124).

What it takes to sign off

• The budget at the right end points. Know whether your channel model runs ball to ball or die to die, and whether the package models are inside it or added by COM.
• Laminate data at 53 GHz, on your stackup. Df keeps rising with frequency, and loss rises with temperature: 1.25 to 1.40 dB/in from room temperature to hot (80 °C) on one measured board (C-119).
• COM, not insertion loss alone. 802.3ck’s 28 dB is a recommendation; the requirement is a channel operating margin of at least 3 dB (C-107, C-108). See compliance masks and COM.
• The package as a budget line. In the contributed 224G examples each package takes 5 to 9 dB of the 40 (C-118).
• The current draft. 802.3dj is still changing; check the latest draft’s numbers before a design depends on them.

Go deeper: rates, the 28 dB curve and the reference receivers

From 100 Gb/s to 53.125 GBd. 802.3ck transcodes 64B/66B blocks into 256B/257B and adds RS(544,514) (C-110). The overhead is 66/64 × 257/264 × 544/514 = 17/16, so 100 Gb/s of data becomes 106.25 Gb/s on the wire, 53.125 GBd at two bits per PAM4 symbol (C-106). The same chain turns 200 Gb/s into 212.5 Gb/s, 106.25 GBd, the rate the 802.3dj baseline uses (C-115). One symbol lasts 18.8 ps at 112G and 9.4 ps at 224G, and Nyquist is half the symbol rate: 26.5625 and 53.125 GHz.

The 28 dB is a curve. 802.3ck’s recommended backplane channel loss is

IL(f) ≤ 0.693 + 2.161 · √f + 0.607 · f   dB,   f in GHz, up to 26.5625 GHz

which reaches 27.95 dB at 26.5625 GHz, the “28 dB” (C-107). It is a constant, a square-root term that behaves like copper loss and a linear term that behaves like dielectric loss: the two-term shape this page uses, plus a constant.

Two test channels. 802.3ck checks a receiver over a low-loss and a high-loss channel, 13.5 to 14.5 dB and 27.5 to 28.5 dB at 26.5625 GHz (C-108). OIF’s CEI-112G-LR uses the same pair, 14 and 28 dB between the package balls (C-123).

Reference receivers. COM scores a channel by running it through a reference receiver. 802.3ck’s is a CTLE and a 12-tap DFE with floating taps, after a 5-tap transmitter FFE (C-108). A feasibility study contributed to the task force for the 40 dB backplane assumed a long feed-forward equaliser and maximum-likelihood sequence detection (MLSE), and said plainly that its settings were not a parameter proposal (C-120). It shows the direction, not the final reference receiver. The equalisation page explains each block.

What the panel leaves out. Loss per inch comes from a first-order line model with the makers’ published Dk and Df: no vias, packages, connectors, crosstalk or reflections, and a narrow differential trace loses somewhat differently. The pulses use the site’s shared channel model, with only the loss at each Nyquist taken from the laminate, and no equalisation. They show the mechanism, not a link margin.

In the real world

Broadcom began shipping Tomahawk 6 in June 2025, a 102.4 Tb/s switch with 200G or 100G PAM4 SerDes, including an option for 1,024 lanes of 100G, and announced a co-packaged optics version (C-121). At 200G the same capacity is 512 lanes. A 1.6 Tb/s port is eight 200 Gb/s lanes (C-112), and every one of them faces the budget on this page. Where the board cannot carry 53 GHz far enough, the lane can go into cable beside the package, into a retimer, or into optics mounted on or beside the package.

Where the numbers come from

112G from the published IEEE Std 802.3ck-2022 and OIF’s CEI-05.3 agreement. 224G from the 802.3dj task force’s public objectives, adopted motions and presentations, not from the draft itself, which is not public; every 224G figure is marked draft in the claim ledger (C-105 to C-124). The comparisons are arithmetic on those figures, checked by the site’s number gate.

Related

• PCIe 6.0 vs PCIe 5.0: PAM4, FLIT, FEC and Signal Integrity
• NRZ vs PAM4 Signaling: Bandwidth and SNR Trade-off
• Compliance masks and COM
• Stackup design and laminate choice
• Chiplet Interconnects: UCIe vs BoW vs AIB Compared