TDEB Top Bar

Trending
Semiconductor Industry News — The Digital Electronics Blog
Follow Us

TDEB Masthead

The Semiconductor Hub

Semiconductors.  Chips and Systems.  Innovation.

Indexed by leading Search Engines, AI and referenced by Engineers worldwide 500+ Articles 10M+ Pageviews 30+ Reports 50+ Citations

Spending 40 dB: A Bump-to-Connector Budget for 224G SerDes

Murugavel Ganesan
by
0
Share Facebook X LinkedIn WhatsApp Email
Flip-chip package on a circuit board with differential traces and twinax cables leading to an edge connector

Take a copper lane from 100 Gb/s to 200 Gb/s and the Nyquist frequency nearly doubles, to 53.125 GHz in IEEE P802.3dj. Yet the loss budget grew. The task force allows 40 dB from die to die, while OIF's 2022 framework lists 28 dB at 28 GHz for CEI-112G long reach. Package, board, connector and receiver all have to share that bigger number at a far harsher frequency.

This post follows one end of the budget, from the die bump to the connector, where the adopted symmetrical case allows 11.5 dB per host. It then prices the ways out. In a worked 7 inch route, a cabled host recovers about 4.4 dB with no extra SerDes, while a retimer on one 8-lane port adds about 17 W under our illustrative 5 pJ/b. Each input is either sourced or marked illustrative, so you can redo the sums.

Key takeaways

  • "224G" is shorthand. P802.3dj signals at 106.25 GBd PAM4, which is 212.5 Gb/s on the lane and a unit interval of about 9.41 ps.
  • The adopted symmetrical copper cable budget is 11.5 + 17 + 11.5 = 40 dB. The package takes about 5 dB of each 11.5 dB host in the task force's example.
  • Board trace costs about 1.4 dB per inch when hot. Twinax in the same example costs about 0.32 dB per inch, which is why cabled hosts matter.
  • Published 224G silicon quotes analog energy, not the whole link. Treat 3 pJ/b as a floor when you size a switch.
  • A retimer doubles the transceivers on a lane. Buy decibels from the channel first where you can.

The idea in plain words

A SerDes lane sends a stream of bits over a pair of copper wires. The faster the bits, the more of the signal the copper soaks up on the way, and engineers count that loss in decibels (dB). Every 20 dB of loss cuts the signal's amplitude by a factor of ten, so a 40 dB channel delivers about one hundredth of the amplitude that left the transmitter, at the frequency that matters most.

The new 200 Gb/s per lane Ethernet work allows 40 dB end to end. That total has to be shared: the chip package and the circuit board at each end take a slice, and the cable in the middle takes the rest. If one end uses more than its share, something else has to give.

When a route is too lossy, designers have three broad choices:

  • Shorten or improve the board path, for example by placing the connector closer to the chip.
  • Run part of the route over cable inside the box, because thin twinax cable loses far less per inch than board traces at these frequencies.
  • Add a retimer, a chip that cleans up and resends the signal. It solves the loss problem but adds power, cost and delay.

The deep dive puts numbers on each choice: where the 40 dB came from, how it splits between package, board and cable, a worked 7 inch route, what the receiver does with the remaining signal, and what a retimer costs in watts.

Deep dive: the numbers and details

This part holds the loss allocations, a worked budget, the receiver and error targets, published silicon and the power model. Skip it if the overview is all you need.

Where the 40 dB came from

The number was a negotiated one. The May 2023 proposal to the IEEE task force (Mellitz et al.) reported SerDes builders asking for 36 to 37 dB and system builders for 42 to 45 dB, and it called 40 dB a compromise. Its COM (channel operating margin) runs reached the 3 dB pass threshold on channels of that loss. The task force adopted 40 dB at 53.125 GHz for backplanes that month, then for copper cable that July.

A later motion split the budget. For the symmetrical cable link, each host gets 11.5 dB and the cable assembly gets 17 dB. Two asymmetrical cases followed, because real hosts are not alike: a 16.5 dB host facing a 6.5 dB host leaves 17 dB for the cable, and an 11.5 dB host facing a 6.5 dB host leaves 22 dB. Each case sums to 40 dB, and the host figures include allowances for the BGA and connector footprint vias.

One caution on status. P802.3dj was still in Standards Association balloting at draft 3.3, with final comment responses dated October 14, 2026, and the IEEE 802.3 ballot page showed no open ballots on November 6, 2026. We did not read the draft itself, and the budget figures here come from task force motions and contributions of 2023 and 2024. Check them against the published clauses. OIF told IEEE in April 2026 that CEI-224G-LR and MR had reached revision 12, with access restricted to IEEE 802.3 participants, so we use no numbers from those documents.

From bump to connector

The clearest public breakdown is in the November 2023 contribution by Tracy et al. These are example allocations for the cable case, not normative limits, and backplanes share the 40 dB without this split. The totals are the budgets. The splits are one way to meet them.

Segment at 53.125 GHzLow host, PCBNominal host, PCBNominal host, cabledHigh host, PCBHigh host, cabled
Package3.9 dB5 dB5 dB9 dB9 dB
BGA transition0.5 dB0.5 dB0.5 dB0.5 dB0.5 dB
PCB trace2.1 dB (1.5 in)6 dB (4.28 in)2.8 dB (2 in)7 dB (5 in)2.8 dB (2 in)
Near-package connectornonenone1 dBnone1 dB
Internal twinaxnonenone2.2 dB (6.8 in, 30 AWG)none3.2 dB (8 in, 32 AWG)
Host total6.5 dB11.5 dB11.5 dB16.5 dB16.5 dB
Routed distance after the BGA1.5 in4.28 in8.8 in5 in10 in

Three things stand out once you divide the columns through.

The trace rate is a hot number. Each PCB entry works out at about 1.4 dB per inch. That matches the August 2023 channel models by Weaver and Chan of Arista, which give about 1.25 dB per inch at room temperature and about 1.40 dB per inch hot, on conventional construction with no skip layers or microvias. Their 9 inch channel loses 1.28 dB more at 80 °C than at 25 °C (15.73 dB against 14.45 dB, with one vendor's connector). Sign off a budget at the hot corner, or that margin is imaginary. Those same models imply about 3.1 dB of fixed loss for the BGA, vias, connector and module trace: 7.32 dB at 3 inches, less 3 × 1.40 dB.

The package varies more than the board does. The task force adopted two reference packages: Class A, optimized for loss, and Class B, for high radix. Feeding their published transmission line parameters into the Annex 93A loss formula gives our own estimate of trace loss alone at 53.125 GHz:

Reference packageLoss per mm (our calculation)Modeled trace lengthsTrace loss
Class A0.153 dB12 mm, 33 mm1.84 dB, 5.05 dB
Class B0.184 dB8, 24, 30, 45 mm1.47, 4.41, 5.52, 8.28 dB

As a check on the method, the same formula applied to OIF's board parameters returns 1.30 dB per inch at 56 GHz, the figure OIF quotes. The results also land close to the 5 dB and 9 dB examples above, once vias and pads are added. The practical message is that shortest and longest lanes in one large package can differ by several decibels, so a single "package loss" entry hides your worst port.

Twinax changes the trade. The 30 AWG example is 2.2 dB over 6.8 inches, about 0.32 dB per inch, more than four times better than the hot PCB rate. For the same 6 dB of host routing, the cabled option covers 8.8 inches and the PCB covers 4.28 inches. OIF's framework goes further and raises cables that terminate on the package itself, skipping the host-to-package transition.

A worked budget: 7 inches to the connector

Suppose a switch has a lane that must travel 7 inches from the package ball to the I/O connector, and the far end is a nominal 11.5 dB host. The route and the choice of scenario are illustrative. The unit rates come from the sources above: 5 dB package, 0.5 dB BGA, 1.4 dB per inch of PCB, 1 dB near-package connector and 2.2 dB per 6.8 inches of 30 AWG twinax.

LineAll PCB2 in PCB + 5 in twinax
Package and BGA5.5 dB5.5 dB
PCB trace7 × 1.4 = 9.8 dB2 × 1.4 = 2.8 dB
Near-package connectornone1 dB
Twinaxnone5 × 0.3235 = 1.6 dB
Bump to connector15.3 dB10.9 dB
Against the 11.5 dB nominal host3.8 dB over0.6 dB under
Left for cable and its 2 connectors (40 less this host, less 11.5)13.2 dB17.6 dB

The cabled route recovers 4.4 dB. The all-PCB route still fits inside a 16.5 dB high host, with 1.2 dB to spare, but the far end must then be gentle or the external cable short. That is the point of the asymmetrical budget: a loss-heavy port is allowed, provided something else gives way.

What the receiver does with what is left

Insertion loss is the easy part to add up. OIF's framework warns that 112G work exposed the effect of return loss in low-loss channels, and that 224G metrics must weigh more than insertion loss. It also shows how a 4 ps intrapair skew distorts a 9 ps pulse. This is why COM, not a dB figure, decides whether a channel passes. Use the dB budget to plan and COM to judge.

The feasibility work suggests how much receiver is implied. The 2023 COM sets (explicitly not a parameter proposal) used a transmit FFE with 4 pre-cursor and 1 post-cursor taps, a receive FFE spanning 6 pre-cursor and either 24 or 60 post-cursor positions, one DFE tap equivalent, and MLSE (sequence detection). That sort of structure is hard to build without digitizing first, which is why the long-reach receiver went digital. If you need the basics, here is a comparison of analog and DSP SerDes design.

Error targets are the other lever. In March 2024 the task force set a detector error ratio (DER0) of 2e-4 for backplane and copper cable links. Comment files on draft 3.3 cite a maximum PMD link BER of 2.761e-4. OIF notes that relaxing a target from 1e-6 to 1e-4 improves SNR margin by 2.19 dB. That is margin at the slicer, not 2.19 dB of channel, and the price is paid in FEC and latency elsewhere.

Published silicon shows the state of the art, with the usual caveat that these are authors' own results on their own test chips:

  • Synopsys, ISSCC 2024 (Pfaff et al.). A 7-bit DAC transmitter and a 7-bit ADC receiver in 3 nm, with up to 20 dB of front-end peaking and 4.0 bits ENOB at 50 GHz. The press kit gives a pre-FEC BER of 5e-7 over a 35 dB bump-to-bump channel at 3 pJ/b analog. The paper's title in the program and in JSSC states 40 dB, and a 2025 invited CICC paper is titled 42 dB post-FEC error free. These numbers are not interchangeable, so read each one's own conditions.
  • Intel, ISSCC 2024 (Cusmai et al.). A transmitter alone: 7-bit DAC, 9-tap FFE, 1 Vppd, 36.0 dB SNDR and 0.92 pJ/b analog, with PAM-4 and PAM-6 modes.
  • ISSCC 2025 paper 7.2 (Mostafa et al.). Titled 2.2 pJ/b at 212.5 Gb/s with more than 46 dB reach in 5 nm. We read only the title record, not the paper.

Do not rank these on a single axis. Node, BER, reach and what counts as "analog" all differ.

The power side of the budget

Here is a simple model for sizing. The 3 pJ/b is the published analog claim, assumed to hold at 212.5 Gb/s. "Analog" suggests that DSP and FEC sit outside that figure, so the second column adds a digital allowance of 2 pJ/b. That 2 pJ/b is illustrative, as are the port and switch scenarios. Substitute your IP vendor's number.

Line3 pJ/b (analog only)5 pJ/b (illustrative total)
One transceiver lane at 212.5 Gb/s0.6375 W1.0625 W
8-lane port, one end5.1 W8.5 W
8-lane link, both ends, no retimer10.2 W17 W
Same link with one retimer per lane20.4 W34 W
Added per port by the retimer10.2 W17 W
512-lane switch, SerDes only326.4 W544 W
Retimers on 16 of its 64 ports163.2 W272 W

The retimer lines follow OIF's point that removing a repeater halves the SerDes on a switch port, from four to two. We assumed a retimer lane burns the same energy as a host lane, and a retimer built for a shorter reach could do better. Even so, the comparison is stark. The 4.4 dB in the worked budget came from routing, with no new SerDes. A retimer removes the budget problem outright, but it adds latency and may not fit the power envelope, a concern OIF raises directly.

The same logic explains near-package and co-packaged optics: a shorter electrical channel means less equalization and no repeater. OIF expects a mix of approaches rather than one winner. For the system side of that argument, see where the bottleneck moves once copper runs short.

Five checks before you freeze the floorplan

  1. Have you budgeted the worst lane? Use the longest package trace and the furthest connector, not an average.
  2. Is the board rate from your stack-up, at temperature? A 0.15 dB per inch gap between room and hot is 1.05 dB over 7 inches.
  3. Which far end did you assume? A high-loss host only works if the partner host and the cable leave room for it.
  4. Did you run COM with real connector and via models? A channel that meets its dB line can still fail on reflections or crosstalk.
  5. Did you quote total energy? Ask for analog, DSP and FEC separately, at your BER and your loss.

Our SoC lifecycle on The Semiconductor Hub shows where package, board and IP decisions fall in a program.

FAQ

Does the 40 dB apply to chip-to-module links?

No. The objective covers backplane and copper cable PHYs. Chip-to-module has its own channels and error allocation. A November 2023 fixture study (Kocsis and Sekel) gives a sense of scale, with a short channel model of 10.32 dB and a mated test fixture of 9.31 dB, of which 2.00 dB is the simulated module connector.

Why not lower the baud rate with PAM6?

OIF's table shows PAM6 at 224 Gbps runs at 90 GBd, with a 45 GHz fundamental instead of 56.25 GHz, at a 3.62 dB SNR penalty. That is a real option in research, and Intel's transmitter supports it, but the task force adopted PAM4 in May 2023, and the KR baseline runs at 106.25 GBd.

Sources

The worked budget and power model use illustrative scenarios and an illustrative digital allowance, as marked. Loss allocations are task force examples rather than limits, and the package loss per mm is our own calculation. Vendor and paper figures are the authors' own claims, as published on the dates shown.

Murugavel Ganesan
About the Author
Murugavel Ganesan
Product Director | SerDes · IVN · Edge AI · Compute | Ex Intel · Infineon

A semiconductor product leader working across product architecture, definition, strategy, and silicon and software execution, from concept through volume production. He has worked with teams across Asia, Europe and the USA, including at Intel, Infineon and ArrayComm.

His work covers ASICs, SoCs, FPGAs and chiplets for wireless, wireline, automotive, IoT and AI, including PCIe, CXL and DDR interfaces alongside RF and mixed-signal design.

Share Facebook X LinkedIn WhatsApp Email

Post a Comment

0 Comments

Your comments will be moderated before it appears here.

Post a Comment (0)

#buttons=(Ok, Go it!) #days=(20)

Our website uses cookies to enhance your experience. Learn more
Ok, Go it!