Blog · verification

Signal integrity on dense boards: one number, four missing parts

An EMS describes simulation, automated test and eye diagrams inside its own plant. When the only document a buyer has is a capability profile, what part of a signal-integrity claim can actually be checked?

ChinaPCB Editorial 2185 words · 10 min read

The switching signals in this article

  1. 01 A tolerance is the last of five parts Object, instrument, frequency, metric and limit. The record publishes the limit and leaves the other four to be inferred.
  2. 02 Fifteen percentages, no object Every line that gives an impedance tolerance omits the coupon it was measured on, which is the difference between a board and a test strip.
  3. 03 The tightest number is conditional Five lines quote ±5%. Four of them attach a condition: a sample order, a 2028 roadmap, a 1.6T product, an inner layer.
  4. 04 Geometry is published, behaviour is not Thirty-five via lines give drill, pad and aspect ratio. None gives a capacitance, an impedance or an antipad diameter.
  5. 05 No dielectric constant anywhere Six profiles name a laminate brand. None gives Dk or Df, which are the two values that set impedance and loss.

The article that prompted this one is written from inside a factory. Its author is an electronics manufacturing services company describing what it does with signal and power integrity: model the board in CAD, run two- and three-dimensional electromagnetic simulation before the first prototype, sweep the finished assembly on automated test equipment, plot the result as an eye diagram, feed the data back into the next layout revision. The frequency plot in its first figure runs from DC to 16 GHz with loss on the vertical axis.

None of that is wrong. It is also not available to a buyer. The test equipment sits behind a factory door, and the eye diagram in the article is an illustration rather than a record. What a buyer can hold before an order is a capability profile, so the question here is narrow: in that profile, which parts of a signal-integrity claim survive?

The counts come from the 130 factory profiles in this directory, 1,252 capability lines in total, every profile at the same evidence level. The PDN article counted the same lines for resistance. This one counts them for the measurement that sits behind the number.

An impedance specification is not one number. It is five, and the number is only usable when all five are present.

Fig 1Five parts of one tolerance, as they appear in 1,252 capability lines

Object

The coupon, its line width, its dielectric

0 of 15 lines

Instrument

TDR rise time, or a VNA sweep

0 of 1,252 lines

Frequency

The point where the figure applies

0 of 15 lines

Metric

Single-ended, differential or common mode

1 of 15 lines

Limit

The window: ±5% to ±10%

13 of 15 lines

A tolerance with no object is a claim about a structure that has not been named. The record supplies the last link and drops the four that make it checkable.

The chain runs left to right in the order a laboratory would work. First you decide what to measure, then with what, then at which frequency, then which quantity, and only then does a window make sense. A profile reverse-engineers the process: it publishes the window and lets the buyer assume the other four.

That order is not pedantry. Two factories can both write ±8% and mean different things. One measured a coupon line on the panel border with a 200 ps TDR edge; the other is quoting a customer’s acceptance criterion from a drawing. The words are identical and the evidence is not.

Fifteen percentages and no statement of what was measured

Thirteen of the fifteen lines that mention impedance carry a figure. Read the same percentage against three different geometries and the window changes size each time.

Quantity the tolerance could apply to±8% windowWidth of the window
A 50 Ω single-ended line46 to 54 Ω8 Ω
A 90 Ω differential pair82.8 to 97.2 Ω14.4 Ω
A 28 Ω power-plane trace at 1 oz25.8 to 30.2 Ω4.4 Ω

Only one line in the directory says which quantity it refers to. Sunly writes “±8% / ±6%” and labels the second figure differential, which is the single case where the metric is stated rather than assumed. Everywhere else the word impedance stands alone, and the buyer supplies the meaning.

The object is missing in all fifteen. Impedance is produced on a test coupon: a strip of line in the panel border, built in the same lamination cycle, measured with a time-domain reflectometer against a stated rise time. The product trace is a different structure. It is narrower after etch, it sits over a plane that has voids in it, and it passes through vias that the coupon does not have.

That difference is the whole reason the measurement exists. A coupon number is meant to prove that the process held its geometry on the panel. Quoted without the coupon, it becomes a promise about a structure nobody described. The directory has no line that mentions a coupon, a test board or a witness strip: zero out of 1,252.

The tightest number is never the standard tier

Five lines quote ±5%, the tightest figure in the directory. Four of the five attach a condition to it.

ProfileHow the ±5% is qualifiedWhat the unconditional figure is
Shennan CircuitsSample order, stated next to a production figure±7% in mass production
Glorysky HuizhouA 2028 roadmap, after ±8% in 2026 and 2027±8% today
Redboard Ji'anOne 1.6T optical-module product, 12 to 16 layersNot stated for other products
Sunshine Global CircuitsInner layers at 1 oz; outer layers are not in the sentence±10%
Sihui FujiNo condition in the line±5% as published

That pattern is not dishonesty. It is how a capability sheet is written: the best number goes on the page with its condition compressed into three words. The problem is what happens downstream. A buyer comparing five profiles sees 5, 5, 5, 5, 5 and reads a market, when three of those fives belong to a sample order, a future year and one product family.

Fig 2The tightest impedance figure each of the fifteen profiles publishes

±5%

5 profiles · 1 without a condition

±6%

1 profile · differential

±7%

2 profiles · both tiered

±8%

5 profiles · 1 qualified

No figure

2 profiles · the word only

the line attaches a condition — the line stands alone

Fifteen profiles, thirteen figures, five of them at ±5%. The tighter the number, the more often it arrives with a qualifier attached, which is what makes a straight comparison of the percentages misleading.

Vias: thirty-five lines of geometry, no electrical value

A via is where a trace stops being a waveguide and becomes a lumped object. Its electrical size is set by four dimensions: the drill, the pad, the antipad, and the depth of the void it leaves in the reference plane.

The directory publishes three of the four, and publishes them as limits: the smallest hole, the smallest pad, the deepest barrel. Thirty-five lines across 26 profiles describe via geometry. Thirty lines give an aspect ratio, from 1:6 on a thin board to 40:1 on a 2028 roadmap. The figures are manufacturing limits: the smallest hole a plant will quote, the deepest barrel it will plate. On an order, the drill is whatever the design requires, not the minimum.

The antipad appears in none of them. That is the dimension that decides how much copper is removed from the plane around the barrel, which is the return-path discontinuity a dense board cannot avoid. The void in the plane is only visible in the thing the design review asks for and the record almost never contains: a cross-section.

Six lines mention back-drilling, and three of those give a residual stub with a tolerance. Read the three together.

Profile and tierStub as publishedBand it actually covers
HannStar Board, standard7 ± 5 mil2 to 12 mil
HannStar Board, advanced6 ± 4 mil2 to 10 mil
HannStar Board, limited5 ± 3 mil2 to 8 mil
Dynamic Huangshi5 ± 3 mil2 to 8 mil
Kingshine, 2.6 mm board± 0.15 mm5.9 mil either side

Three tiers, one lower bound. Every HannStar band starts at 2 mil and the tiers differ only at the top, which means the “advanced” grade is a promise about the worst panel rather than about the typical one. Kingshine uses a different grammar altogether: a single drilling tolerance of ±0.15 mm, which is 5.9 mil in each direction on a 2.6 mm board.

At 16 GHz, the frequency the source article plots, none of this is the dominant term. In a laminate with a dielectric constant of 4, the wavelength is 9.4 mm, and a 5 mil stub is 0.127 mm, which is 1.4% of a wavelength. A residual that short is electrically small. The number that decides the via’s behaviour at that frequency is the barrel’s own capacitance against the plane void, and that number is published nowhere in the directory.

Six profiles, eight laminate brands, no dielectric constant

Rows in the directory name Rogers, Megtron, Ventec, Isola, Arlon, Taconic, Shengyi and Kingboard, and several name PTFE as a family. Six profiles name at least one of these brands, and one of them names four in a single line.

Not one line gives a dielectric constant or a loss tangent. The closest thing to a value is a grade: “Very Low Loss or better” on a Delton switch board, which is a purchase category rather than a number. A grade tells a buyer what the plant is willing to buy. It does not tell the design what the impedance will be, because Dk sets impedance at a given geometry and varies within a grade with resin content and glass style.

The datasheet does carry a number, and that number has a frequency attached. A laminate sold as Dk 3.9 at 1 GHz is not 3.9 at 16 GHz, and the shift is not the same for every grade. That is why the source article’s insistence on material consistency is correct and why a brand name cannot satisfy it: consistency is a property of a lot, and the record speaks in families.

Three frequencies and no measurements at any of them

Three lines mention a frequency at all. Sihui Fuji publishes a supported range of 1 GHz to 40 GHz. Eleceltek Yangzhou lists 77 GHz and 79 GHz radar boards. Dynamic Huangshi publishes a high-frequency laminate extended to 90 GHz.

Fig 3Where the record makes a frequency claim, and where it makes none
0255075100 GHz
The source article's sweep
DC to 16 GHz, measured
Sihui Fuji
1 to 40 GHz, supported
Eleceltek Yangzhou
77 and 79 GHz radar
Dynamic Huangshi
90 GHz laminate
Insertion loss, return loss, TDR, eye diagram
no figure at any frequency
0 of 1,252 lines

Three claims about capability, one measured sweep that lives inside a factory. Nothing in the directory is a measurement, and the words that would identify one — insertion loss, return loss, TDR, eye height — appear nowhere in 1,252 lines.

The first three are statements about what a plant can build. They are not measurements, and they are not meant to be: a radar board sold to an automotive tier-one is qualified by the customer’s own test, not by the fabricator’s capability page.

The absence is still the finding. The words that would identify a measurement do not occur. Insertion loss: zero. Return loss: zero. Time-domain reflectometry: zero. S-parameters: zero. Eye diagram, jitter, skew, length matching, crosstalk: zero each. The only electrical quantity in the entire directory is a percentage tolerance, and it arrives without the four parts that would make it a measurement.

What a buyer can put in the RFQ

The source article ends with a factory that measures everything and feeds the result back into design. That loop is real, and it belongs to whoever owns the test equipment. A buyer without access to the loop can still ask for the four missing parts, and the questions are short enough to put in writing.

  1. Name the coupon. Did the impedance figure come from a coupon on my panel or from a separate test board, and what line width and dielectric thickness does that coupon carry?
  2. Name the instrument and the edge. A TDR with what rise time, or a sweep to what frequency? A tolerance measured with a 200 ps edge and one measured with a 35 ps edge are not the same claim.
  3. Name the quantity. Single-ended, differential or both, and on which layer. Sunly is the only profile in the directory that answers this in the same line.
  4. Name the via. Drill, pad and antipad actually used on this order, and whether the test coupon carries a via structure as well as a line.
  5. Name the lot, not the brand. Dk and Df at the frequency the design runs at, taken from the laminate lot's datasheet rather than the family name on the capability page.

None of those five is exotic and all of them are answerable in a single email by a plant that knows its process. The RFQ checklist covers the rest of the order; these five cover the part of it that a capability profile structurally cannot.

The source article’s most useful sentence is buried in the middle: signal integrity can be degraded by hidden fabrication defects. That is exactly right, and it is the reason a percentage without an object is worth so little. A defect in impedance is invisible until something measures it against a structure that was defined in advance. The record sells the limit. The structure is the part the buyer has to ask for.

Source: capability lines published by the 131 manufacturers profiled in this directory, read 2026-11-17. Coupon practice, TDR rise time and the Dk frequency dependence are standard test practice, not profile figures.

Questions this article answers

Why does an impedance tolerance need a test coupon named next to it?

Because the number is produced on a test structure, not on the product trace. A coupon has its own line width, dielectric thickness and reference plane, and the tolerance describes that structure. Applied to a product trace, it assumes the etch and the laminate are identical on both, which is the assumption the measurement was meant to remove.

Is a ±5% impedance tolerance better than ±8%?

It is tighter, and four of the five profiles quoting ±5% in this directory attach a condition to it: a sample order, a 2028 roadmap, a 1.6T optical product or an inner layer. A tolerance is only comparable when the object, the instrument and the frequency are the same, and none of those three is stated in any of the fifteen lines.

What should an impedance test coupon carry?

A line of the same width and copper weight as the product trace, on the same layer, with the same reference plane, plus the dielectric thickness it was built on. Without those four, the measured impedance belongs to a different geometry. Ask whether the coupon came from your panel or from a separate test board, because those are different levels of evidence.

Do vias need their own impedance specification on a dense board?

A via is a lumped discontinuity, and its size is set by the drill, the pad, the antipad and the void left in the reference plane. Capability lines publish minimum drill and pad as manufacturing limits, which is the smallest hole a plant will quote, not the hole used on an order. If the stack has more than a few hundred microns of stub, ask for the residual as a measured value rather than a nominal.

Why do factory capability sheets avoid Dk and Df?

Because a laminate grade is a purchase category and Dk is a property of the lot. The same grade varies with resin content and glass style, and the datasheet value is quoted at one frequency while the design runs at another. Naming Rogers, Megtron or Taconic describes what the plant can buy, not what the panel will measure.