MCIO / SlimSAS Cable Length & Insertion Loss Calculator

How it is calculated

The channel is summed segment by segment at the generation's Nyquist frequency, the way the specifications budget it, and compared with the total the generation allows. Losses in decibels add, so the sum is a plain addition; the work is in the figures that go into it.

channel loss = host package
             + trace length × board loss per inch
             + vias, caps & breakout
             + mated pairs × loss per pair
             + cable length × cable loss per cm
             + far-side board (+ slot connector, on a riser)
             + end device, or the retimer's package

margin       = budget − channel loss
verdict      = pass      if margin ≥ 10 % of the budget
               marginal  if 0 ≤ margin < 10 %
               fail      if margin < 0

longest cable, recommended = (0.9 × budget − everything but the cable) ÷ cable loss per cm
longest cable, absolute    = (budget − everything but the cable) ÷ cable loss per cm

With a retimer the budget restarts on each side of it, so the sum covers only the side the cable is on: from the retimer's package to the device when it sits on the host board, from the host package to the retimer's package when it sits on the far board. The 10 % design margin is the low end of the 10 to 20 % that practice keeps for temperature, humidity, reflections, crosstalk and supply noise [15].

Default values, and where each comes from

Every default below can be overwritten in the form. The numbers in brackets point to the sources at the end of the page.

What the presets rest on

30 AWG cable

Samtec's 85 Ω 30 AWG FEP twinax datasheet curve reads about 6 dB per metre at 16 GHz and 3.9 dB at 8 GHz; the DesignCon 2023 measurements of Gen5 MCIO cables average about 0.06 dB/cm. A 2017 3M catalogue puts its 30 AWG at 7.2 dB per metre at 12.89 GHz, so a given cable can sit a third higher. [8][2][16]

32 AWG cable

Samtec's 32 AWG datasheet curve reads about 6.5 dB per metre at 16 GHz and 4.4 dB at 8 GHz, for a 100 Ω cable; an 85 Ω PCIe cable of the same gauge loses a little more, so the preset is rounded up. [9][7]

34 AWG cable

Samtec states 0.20 dB per inch at 16 GHz for 34 gauge twinax (0.079 dB/cm); its PCIe 6.0 article gives 0.15 to 0.20 dB per inch depending on gauge, and PCI-SIG's cabling talk uses about 0.19 dB per inch. The 8 GHz value is scaled from 16 GHz. [7][10][4]

Ultra-low-loss board

PCI-SIG's cabling talk uses about 1 dB per inch at 16 GHz for ultra-low-loss board; the DesignCon 2023 paper's reference for Tachyon 100G stripline is 0.85 dB per inch. [4][2]

Low-loss board

Samtec budgets 1.1 dB per inch at 16 GHz on low-loss material; the DesignCon 2023 paper measured 1.2 to 1.5 dB per inch on stripline with a loss tangent of 0.002 to 0.004. [7][2]

Mid-loss board

Samtec's PCIe 6.0 article puts mid-range dielectrics as high as 1.4 dB per inch at 16 GHz; the DesignCon 2023 paper's reference for FR408HR stripline is 1.33. [10][2]

Standard FR-4 board

A rule of thumb, not a measurement: standard FR-4 is a Gen1 to Gen3 material and no source quotes it at 16 GHz; the DesignCon paper's microstrip measurements on mid-loss board ran 2 to 4 dB per inch. [2]

Gen4 board figures

The 8 GHz presets follow the 16 GHz figures scaled for frequency; Samtec's own Gen4 example uses 1.35 dB per inch for mid-range material, the conservative end. [7]

MCIO connectors

SFF-TA-1016 limits a mated connector pair to 1 dB up to 16 GHz for every line rate it lists, 32 GT/s included. Measured harnesses agree: the DesignCon 2023 paper's 1 m cable with both connectors came to 8 dB against 6 dB of bare cable, and NVIDIA's 250 mm MCIO harness measures 3.5 dB with its connectors. [5][2][11]

SlimSAS connectors

SFF-8654 is a mechanical specification: its electrical speed requirements were removed in Rev 1.0 and it names SAS as its application, so there is no published loss limit. The tool assumes the MCIO figure; replace it with your connector maker's data. [6][20]

PCIe 4.0 split

Samtec's 2021 example budgets the Gen4 connector at 0.5 dB; the tool keeps PCI-SIG's 1.5 dB. The non-root package figure is the Gen5 reference, used as a conservative stand-in. [1][7]

PCIe 6.0 split

The paper splits 32 dB into 8.5 dB for the add-in card (4.6 dB of it package) and about 15.7 dB for the host board traces. The 7.8 dB left is the root package and the slot connector together; the tool keeps the connector at 1.5 dB and gives the package the rest. [3][12]

A figure often quoted that is not used

NVIDIA's BlueField-3 guide gives 0.24 dB/cm for a harness with its connectors. That harness is a Cabline CA-II Plus micro-coax assembly, not an MCIO twinax cable, and its own table works out to about 0.19 dB/cm plus a dB for the two ends. NVIDIA's MCIO figure is the ConnectX-8 one: 3.5 dB for a 250 mm harness with both connectors, which the defaults here reproduce within half a decibel. [13][11]

Disclaimer. Estimates only. Insertion loss is one of several things that decide whether a link trains and holds its rate; return loss, crosstalk, skew and equalization are not modeled, and real cables vary by lot and by vendor. Validate a design with S-parameter measurement of the actual assembly and channel, and treat this as a budget check, not a compliance result.