Blog · verification

PDN resistance: the half of R that the fab decides

Which part of a power delivery network's resistance can a layout engineer fix in CAD, and which part is decided by copper thickness, plating and registration at the factory?

ChinaPCB Editorial 2738 words · 13 min read

The switching signals in this article

  1. 01 Four inputs, one of them yours R = ρL/A has four quantities behind it. The layout file fixes the drawn ratio. Resistivity, plated thickness and drilled length belong to the plant.
  2. 02 27 ceilings, no floor Twenty-seven capability lines give a copper thickness with a unit, across 25 profiles. Not one states a minimum finished thickness, a tolerance, or a per-panel floor.
  3. 03 One word, two quantities Thirteen profiles publish an impedance tolerance. All thirteen describe signal traces. A power plane's DC resistance is a different number by three orders of magnitude.
  4. 04 The barrel nobody publishes No capability line in 1,252 states hole-wall copper thickness. The single mention is a gauge error, not a spec.

An article that circulates widely on power delivery networks opens with the circulatory system and then spends its length on seven techniques: solid planes, thicker copper, via arrays, more power layers, shorter paths, better conductors, a simulation pass before release. Every one of those is a CAD action. That is not a fault in the advice. It is a description of one end of the problem.

The same article prints the equation that shows why one end is not enough. R = ρL/A. Four quantities stand behind the resistance of a power path, and the layout file fixes exactly one of them: the ratio of length to cross-section that you drew. The other three are settled on a shop floor. Resistivity belongs to the foil the plant buys and the roughness left on it after bonding. Length belongs to drill depth, back-drill stub and layer registration. Cross-section belongs to how much copper survives plating and etching, which is not the same number as the foil weight on your drawing.

This piece asks a narrower question than a design guide. Of those four inputs, which ones can a buyer actually find a number for in a factory’s published record, and what does the number measure. The counts come from 131 factory profiles in this directory, 1,252 capability lines in total, every profile at the same evidence level.

R has four inputs and the record covers one

Start with the inventory. Read the equation as a specification and ask, for each term, where the number would come from.

InputWhat decides it at the plantCapability lines with a figure
ρ, resistivity of the conductorFoil grade, purity, and the surface profile left after bonding. Roughness raises effective resistance at frequency and changes the copper's cross-section at the interface.0 lines mention roughness or copper profile.
T, thickness of copperBase foil weight, plating current and time, then etch loss. The shipped thickness is not the ordered weight.27 lines give a thickness with a unit. None gives a floor or a tolerance.
W, width of the conductorImaging and etch. The finished width is narrower than the drawn width, and the loss is a function of copper weight.1 line gives an etch tolerance: ±15%, or ±10% at best capability.
L, length of the pathLayer registration, drill depth, back-drill stub, and where the plane actually ends.2 lines give layer registration. 3 lines give a back-drill stub figure.

The asymmetry in that column is the whole argument. A designer spends a week on the ratio and the plant owns three of its four terms, yet only one of those three arrives with a number attached, and the number is an upper bound. The other two arrive as prose: a phrase like “impedance control” with no window, or “back drilling” with no stub.

There is a second asymmetry. A capability table is written by sales and engineering for the purpose of winning work, so it answers the question “how much can we do”. A purchaser needs the answer to a different question, “how little will you give me on a bad day”, and that question has no column.

Copper thickness is published as a ceiling

Twenty-seven capability lines across 25 profiles give a copper thickness with a unit attached. Grouped by grammar, they say six different things.

ProfileWhat the line saysWhat kind of number it is
Shennan Circuits"6 oz / 30 oz", mass production and sampleTwo ceilings, one per order type
Sunly PCB"6 oz / 12 oz", standard and advancedTwo tiers of ceiling
Glorysky Huizhou6 oz in 2026, 8 oz in 2027, 9 oz in 2028A roadmap, dated by year
HannStar Board JiangyinInner 0.5/1/2 oz, advanced 3, limited 4; outer 0.5/1, advanced 2, limited 3Four tiers, split by layer
Sihui Fuji"3 oz to 15 oz", with a UL-recognised 15 oz on two layersA range plus a certification ceiling
Jinhong Ji'anOuter 1–6 oz (35–210 µm), "1 oz default"; inner 0.5 oz (17 µm)A range, a conversion, and the only default in the directory

Six grammars, one meaning: this is the most we can plate. The number a buyer wants is the other direction — finished copper on the 1 oz layers will be at least 28 µm on every panel of this order, checked at the panel corner as well as the centre — and that sentence appears zero times in 1,252 lines.

The heaviest ceilings in the set make the point sharper. One profile publishes 12 oz and calls it an Asian first. Another publishes 30 oz as a sample ceiling. A third offers 1, 2, 3 and 4 oz as an order-entry menu. All three are selling capability at the top of the range, which is a legitimate thing to sell. None of them is telling you what happens at the value almost every board actually uses. Only one line in the directory names a default, and it names 1 oz.

What a 1 oz design actually ships with

Three multipliers sit between the resistance in your spreadsheet and the resistance in the board, and all three push the same way.

Effect on a 1 oz layerMultiplier
Finished thickness at the Class 2 floor, 80% of nominal: 28 µm instead of 35 µm1.25
Etch and plating loss on a nominal 35 µm foil, landing at 30–32 µm1.09 to 1.17
Junction temperature at 85 °C instead of 25 °C, copper at 0.00393 per °C1.236

Take the first and the third together, because they are independent. A layer that is 28 µm thick and running at 85 °C carries 1.545 times the resistance of the 35 µm at 25 °C that the calculation used. On a rail with a 50 mV budget, that is the difference between passing and failing, and nothing in the two numbers is unusual: 80% is the Class 2 floor for finished copper, and 85 °C is a normal case temperature inside an enclosure.

This is arithmetic on published limits, not a measurement of any specific factory. It is offered because it sets the size of the question. If a design has 10% of headroom against its IR-drop budget and the thickness term alone can move by 25%, then the tolerance on that term is a design input, and a buyer who does not have it is designing against a nominal number that no plant promised to hit.

The copper weight also interacts with width in a way that is easy to miss. Etch loss is a fixed depth, so the same etch that removes 3 µm from a 70 µm layer removes proportionally twice as much from a 35 µm layer. Thin copper and fine lines are the same problem seen twice, which is why the profiles that publish 2.8/2.8 mil minimum spacing attach the figure to 1 oz rather than to a bare line width.

Fifteen impedance tolerances, none of them on a power plane

Impedance is the one electrical quantity the directory does publish, and 13 of the 15 lines that mention it carry a percentage. Read the set.

ProfilePublished impedance toleranceWhat it is attached to
Sihui Fuji±5%Impedance-controlled signal layers
Shennan Circuits±7% mass production, ±5% sampleControlled impedance, quoted with a coupon test
Kinwong±7%High-layer-count signal boards
Delton Guangzhou±8% "on some nets"A switch-board design envelope
HannStar Board Jiangyin±10%, with ±8% inner and ±10% outer for advanced workSignal layers, tiered by product class
Sunshine Global Circuits±10% / ±5%Inner-layer traces "at 1 oz"

The last row is the closest any profile comes to connecting the two ideas, and it still does not connect them: the copper weight is named as the context of the trace geometry, not as a power plane.

The quantities are not comparable. At 50 Ω, ±8% is a window of 4 Ω. At 3 mΩ, ±8% is 0.24 mΩ, and the instruments and coupons that hold the first number have nothing to say about the second. A process that holds characteristic impedance to ±8% proves that it controls etch, dielectric thickness and line width together to a known band. That is real capability and it is worth having. It is not a promise about DC drop, and a buyer who reads the ±8% as a promise about the power net has misread which quantity was measured.

There is one line in the directory that shows how the ambiguity gets into a quote. A design envelope lists “very low loss laminate or better, 0.09/0.09 mm line and space, ±8% impedance on some nets”. Three clauses, three different scopes, and the qualifier “on some nets” is doing work that no acceptance test can reproduce. If the ±8% applies to the nets the supplier chose, then the specification has already been negotiated by the supplier.

A via is a resistor only if the barrel is plated

The design guide offers one worked example: a power via rated at 0.75 A, a rail at 2 A, therefore at least three vias per layer transition, and parallel resistances divide, so R/3. The arithmetic is right. The premise is the part that needs a record.

No capability line in 1,252 states hole-wall copper thickness. The only mention of the barrel at all is a gauge: one profile publishes a hole-wall measurement error of 1 to 2 µm on blind and buried boards, which describes the instrument, not the specification. A second profile states a plated through-hole aspect ratio ceiling of 20:1, and a third a blind-via plating aspect ratio of 0.8:1. Both are geometry. Neither says how much copper ends up on the wall.

That omission matters because the number is not stable. Throwing power falls with aspect ratio and with hole diameter, so an 18 µm wall and a 25 µm wall are both inside what a plant will ship, and they differ by 39% in resistance — inversely proportional to thickness, and a via’s resistance is dominated by the barrel, not the pad. Three barrels at 18 µm are not R/3 of one barrel at 25 µm. They are about R/2.2 of it, which is a different claim from the one the design guide makes.

The practical consequence is a question worth asking in writing, because the answer changes the arithmetic: what is the minimum hole-wall copper on the smallest power via in my stackup, and how is it checked. A plant that answers in micrometres with a coupon and a sampling rule has given you something you can multiply. A plant that answers “IPC Class 2” has given you a class, and the class permits a window.

Length is the one input that does get numbers

Layer registration appears twice, and both times with a figure: ±50 µm at one plant, 50 µm at another. Back drilling appears six times, and three of those give a stub figure — 5 ± 3 mil, a tiered 7 ± 5 down to 5 ± 3 mil, and ±0.15 mm on a 2.6 mm board.

Look at what the first figure means. A 5 ± 3 mil target is a window from 2 to 8 mil, so the worst case is 60% longer than nominal. On a 3 mm backplane with a 20-layer stackup, that stub sits in series with the power path, and it is the one segment of the path where a tolerance is published at all. The paths in a PDN are not a single trace. They are plane, then a via field, then plane, then a trace to the pin, and the published tolerances cluster at the transitions rather than along the runs.

This is the useful inversion in the whole picture. The input that designers think about least, the drilled length, is the one where a factory will put a number on paper, because drilling accuracy is a machine specification and machines come with datasheets. The input that matters most, the copper the current flows through, is a process outcome, and process outcomes get published as capability rather than as tolerance.

Three records that come nearest to a measurement

There are profiles that describe inspection in enough detail to be worth reading, and it is worth being precise about what each one can prove.

One plant retains a cross-section per lot, held for 10 years on standard product and 15 on automotive. That is the only arrangement in the directory that could answer a question about copper thickness after the board has shipped, because the coupon is physical, dated and kept. It proves the lot was sectioned. It does not by itself prove the plane in your board was at the top of the window, because a coupon is cut from a test area rather than from the power layer you care about.

A second plant describes 100% electrical test, AOI at key stages, microsection and four-wire test. The four-wire method is the right instrument for milliohm work, and it is the only technique in the directory that could measure a power net’s resistance directly rather than inferring it from geometry. What is not published is a limit: no figure, no net, no acceptance criterion.

The third describes open and short, impedance, solderability, thermal shock and microsection. A fourth publishes an electrical-test resolution at the milliohm level. Put the four together and you have: copper that can be sectioned, a method that can resolve milliohms, and no profile in the directory that joins them into a number a purchaser can accept or reject.

Four lines that can go into an RFQ

The gap is narrow enough to close in one email. These are written to be pasted, with the placeholder left in.

  1. Finished copper thickness on the 1 oz power layers: minimum 28 µm, measured on a coupon from the panel corner as well as the centre, one section per lot, reported with the shipment.
  2. Minimum hole-wall copper on power vias down to 0.3 mm diameter: 20 µm, with the measurement method named and the sampling frequency stated.
  3. Layer-to-layer registration: ±50 µm or better, declared on the fabrication drawing rather than left to the plant's default.
  4. DC resistance of the 1.2 V rail from the regulator output pad to the load pin: reported once at first article using a four-wire measurement, at 25 °C and at operating temperature.

Each line turns a design assumption into something a plant can accept, decline or counter. The fourth is the one most likely to be declined, and a decline is informative: it tells you the plant tests connectivity rather than resistance, which is the normal case and is worth knowing before the boards arrive.

If a quote comes back with a percentage where a floor was requested, that is the failure mode this whole piece is about. A percentage of nominal is not a floor, and it moves the question back to the buyer in a form that looks like an answer.

How the counts should be read

Every profile in this directory sits at one evidence level, public-sources-reviewed: the capability table was read from the factory’s own published pages, and none of the 131 has been independently audited by us. So a count of zero means no figure is published, not that no plant can do the thing. Plenty of plants measure copper thickness every day; the measurement lives in a QC record that never reaches a website.

What makes the zeros worth counting anyway is that they are uniform. Across 131 profiles and 1,252 lines, the pattern is not scattered — it is systematic. Capability is published as a ceiling, process control is published as a class, and the two numbers a purchaser needs in order to compute a resistance are absent from both. That is a fact about how the industry sells, and it is the reason the four lines above are worth sending.

A power delivery network fails at the point where the design assumption and the shipped copper disagree, and that point is always a document. Ours is a directory of what those documents say. The rest of the work is to notice which sentence is missing.

Source: capability tables and testing descriptions published by the 131 manufacturers profiled in this directory, read 2026-11-17. Copper temperature coefficient and the Class 2 thickness floor are standard values, not profile figures.

Questions this article answers

Does specifying 2 oz copper guarantee a lower PDN resistance than 1 oz?

It lowers the nominal resistance and it raises what the plant has to hold. The number on the order is a nominal foil weight, while the copper a current actually flows through is the finished thickness after plating and etching. At the low end of the permitted window, a 1 oz layer is 28 µm, not 35 µm, and the resistance of that layer is 25 percent higher than the figure in your calculation.

What copper thickness tolerance should I ask for on a power layer?

Ask for a minimum finished thickness in micrometres and a sampling rule, not a percentage. A percentage of nominal leaves the floor open at the bottom of the window, and it says nothing about where on the panel the measurement was taken. The plating current distribution across a panel is what decides whether the corner of the panel matches the centre, and a single coupon cut from the middle will not show it.

Can a factory measure the DC resistance of my power net before shipping?

Two profiles in this directory describe four-wire or Kelvin test equipment, and one of those states a resolution at the milliohm level. Neither is offered as a PDN acceptance test with a limit per net. What a factory normally tests is continuity and isolation, which proves the net is connected, not that it is low enough in resistance.

Why do impedance tolerances of ±5% to ±10% not help with IR drop?

Those figures apply to controlled-impedance signal traces, where the quantity is characteristic impedance in the tens of ohms and the tolerance is a window of several ohms. A power plane's DC resistance sits in the milliohm range, so the same percentage is a window of fractions of a milliohm. A process that holds 50 Ω to ±8% is not thereby holding 3 mΩ to ±8%.

Is a thicker copper layer always the right answer for a high-current rail?

Only if the bottleneck is the plane. In most dense boards the path is plane, then a via array, then another plane, then a trace to the pin, and the via array is where the cross-section collapses. Adding copper weight to the planes while leaving eight 0.3 mm vias between them moves the resistance by very little.