Menu

PCB Design for Manufacture and Test

PCB Design for Manufacture and Test: The Board You Can Actually Build

Notes · Circuit boards and EMCNovember 19, 202614 min readSenova Systems Research

The average PCB project needs 2.9 respins at $46,000 each, and teams that simulate first need 3.0 spins to production against 4.6 for build-and-test teams. A board that works on a bench is not the same as a board a factory can build, test, and repair. The rules that close that gap, in plain words.

First, a simple picture

A recipe and a restaurant kitchen.

Your recipe works in your own kitchen. You make the dish, it tastes right, everyone is happy. Then you hand the recipe to a restaurant that must make it two hundred times a night, every night, with different cooks, in a kitchen you have never seen. The recipe says "a pinch of salt", the restaurant needs grams. It says "cook until done", the restaurant needs a temperature and a time. It says "use the good pan", the restaurant has forty pans, and none of them is yours.

A printed circuit board (PCB) that works on your bench is the recipe in your own kitchen. A PCB that survives a production run is the recipe in the restaurant. Design for manufacture (DFM) is writing the recipe so the restaurant can cook it: every measurement exact, every step the machines can actually do. Design for test (DFT) is writing the recipe so the restaurant can check each dish before it leaves the kitchen, without opening the plate.

This article is about the gap between the two, and the rules that close it.

What a board actually is

A PCB is a sandwich. The bread is fibreglass (usually FR-4). The filling is copper: thin sheets etched into traces, the wires, on several layers, connected between layers by vias, which are small plated holes. Pads are the flat copper areas where parts are soldered. The solder mask is the green (or black, or white) coating that covers everything except the pads, so solder only sticks where it should. Silkscreen is the white printing: part names, pin 1 marks, warnings.

The factory's job has three parts: fabrication (make the bare board: etch the copper, drill the holes, apply the mask), assembly (place the parts and solder them, a pick-and-place machine puts the parts down, a reflow oven melts the solder), and test (prove the board works before it ships). DFM is the design's half of fabrication and assembly. DFT is the design's half of test.

A board that is easy to fabricate, easy to assemble, and easy to test is a board the factory can build at volume, at yield, at a price that holds.

The cost of getting it wrong, measured

The numbers are not small, and they are the reason this article exists.

The average PCB design project needs 2.9 respins. That is Lifecycle Insights' research, cited by Design007. A respin is a new version of the board, new fabrication, new parts, new test, because the last version had a fault. Another summary of the same research puts the average at 2.8 respins at $46,000 each for production-grade boards. A simple prototype respin costs about a week to ten days, mostly waiting for boards to be made and shipped again. A complex production respin costs about four weeks and $10,000–$50,000, once engineering time, lab time, and the delay are counted.

Simulating first cuts the count. A 2026 write-up on the cost of design spins puts teams that find problems by building and testing at an average of 4.6 spins to production, and teams that simulate first at 3.0. That is the same lesson as Article #3 (tune the loop in simulation before anything moves), wearing a PCB badge: the bad versions happen on a screen, not in a factory.

Changes after the design freeze are the expensive ones. The Accuris / Fuld survey (439 industry professionals, March 2026) found 68% make more than six component changes per design project, 40% experience over 20% design rework, and 51% report that over 11% of their designs require component changes after the design freeze, the point where specifications are locked and tooling begins. 46% estimate the average cost of a single post-freeze change exceeds $50,000, and for complex assemblies in aerospace, defence, and automotive it can reach $250,000 per change event. Engineering rework alone runs $15,000–$80,000 per change event, and a single forced redesign totals $135,000–$930,000+ in those industries.

The factory's own numbers agree. On HDI boards (high-density interconnect, the fine, dense boards in phones and wearables), a DFM issue discovered during fabrication costs a full re-order, new NRE, and lost schedule, respins average $3,500–$9,000 in direct and schedule cost. NRE (non-recurring engineering: CAM setup, drill programs, first-article inspection) runs $600–$1,800 per design, and it is charged again on every revision. The factory absorbs 3–8% scrap internally, that is the yield loss built into your quote.

What a board respin costs, and how many to expect (Lifecycle Insights; Tovan 2026; Accuris and Fuld survey, n = 439)

MeasureFigure
Average respins per project2.9
Average cost of one respin (production-grade boards)$46,000
Spins to production: build-and-test against simulate-first teams4.6 against 3.0
Average change after design freezeOver $50,000 (up to $250,000 in aerospace, defence, automotive)

DFM: the rules that let the factory build it

The rules below come from IPC-2221 (the generic standard on printed-board design) and IPC-6012 (the performance specification for rigid boards), as published by the fabricator guides that implement them. They are minimums, not targets, and the factory's own capabilities are the final word. IPC Class 2 is standard commercial electronics; IPC Class 3 is high-reliability (medical, industrial, aerospace).

Traces and spacing, the wires. A trace is a copper wire on the board. The minimum trace width is 5 mil for Class 2 (a mil is a thousandth of an inch, 0.0254 mm) and 4 mil for Class 3, but the recommended width is 6–8 mil, because a trace at the minimum is a trace at the edge of what the etching process can do reliably. The same for spacing: 5 mil minimum, 6–8 mil recommended, and 10 mil improves yield further. Wider spacing is also a voltage rule: IPC-2221 requires 0.1 mm for 0–30 V, 0.4 mm for 150–300 V, and 1.5 mm for 400–500 V between conductors. For power, a 1-amp trace on 1-ounce copper needs 20 mil at a 10 °C rise; 3 amps needs 60 mil. Use 45° corners, not 90°, sharp corners concentrate electric fields and trap etchant (acid traps) during fabrication, where the etching chemical pools and eats the copper.

Vias and annular rings, the holes. A via is a plated hole that connects one layer to another. The minimum drill is 8 mil for Class 2 and 6 mil for Class 3; the recommended is 10–12 mil. The annular ring is the copper ring around the hole: minimum 5 mil, preferred 7–10 mil, because the drill wanders, manufacturers typically allow ±0.075 mm (3 mil) of drill registration error, and a ring thinner than the wander breaks the connection. The aspect ratio (board thickness to drill diameter) must stay under 8:1 for Class 2 and 10:1 for Class 3, or the plating inside the hole gets unreliable; 6:1 is the recommended working rule.

Placement, the parts. The pick-and-place machine needs room: 40 mil minimum between components for its clearance, 100 mil around tall parts so they don't collide, 200 mil between heat-generating parts so they don't cook each other, and 100 mil around mounting holes for screw heads and washers. Keep traces 15–20 mil from the board edge (10 mil is the minimum) so the board doesn't break during profiling and handling. Put fiducials on the board, small copper marks the machine's camera reads to know exactly where the board is. Without them, the machine is placing blind.

The component problem, in one line. The Accuris survey says 68% of professionals make more than six component changes per project, and semiconductor lead times nearly doubled in a single month between February and March 2026, with top-component lead times reaching 40 weeks. A part that is in stock today may not be in stock when the boards arrive. The DFM answer is to design in an alternate part from the start, a second footprint, a second source, before the freeze, not after it.

DFT: the rules that let anyone test it

A board that cannot be tested is a board whose faults reach the customer. Test coverage is the percentage of possible defects your test can detect, and it decides the defect escape rate, the bad boards per million that ship.

What the test can see decides what customers get: defective boards that escape, per million shipped (AtlasPCB DFT guide; IPC-9252A, IEEE 1149.1)

TestFault coverageEscapes per million
Visual inspection only40 to 60%500+
Flying probe85 to 92%50 to 100
In-circuit test (ICT)95 to 98%10 to 30
ICT plus JTAG boundary scan98 to 99.5%2 to 10
Plus functional test99.5 to 99.9%Under 5

Test points, the pads a probe can touch. A test point is an exposed copper pad on a net, placed so a probe can touch it. Per IPC-9252A, the pad must be at least 0.9 mm (35 mil) across, 1.0 mm (40 mil) recommended, with 1.27 mm (50 mil) minimum centre-to-centre pitch and 2.54 mm (100 mil) recommended. Put them on the bottom side of the board, so a bed-of-nails fixture, a board full of spring probes, can touch them all at once while the parts sit undisturbed on top. Never put a test point under an IC: a probe cannot reach it, and the point is wasted. The target is ≥95% net accessibility, every signal net, power rail, and ground with at least one test point, and ≥98% for Class 3. The 5% allowance covers nets that physically cannot reach a pad, like BGA inner balls.

In-circuit test (ICT), the bed of nails. ICT tests 200–400 nets at once in 15–90 seconds per board, and it catches shorts, opens, wrong values, and wrong parts. It needs the fixture, $10,000–$30,000 per board revision, and the test points on a 50- or 100-mil grid so the probes align. That fixture cost is the reason ICT is a volume decision: it pays for itself above a thousand boards a month, and it hurts below it.

Flying probe, the slow, flexible alternative. A flying probe tester moves two or four probes across the board, one net at a time. No fixture, no grid, but 3–10 minutes a board, and 85–92% coverage. It is the right tool for prototypes and runs under 500 a month, and the wrong tool for volume.

Boundary scan (JTAG), the test that needs no probe. IEEE 1149.1 (JTAG) puts a tiny shift register at every pin of a compliant chip, so the test can drive and read the pins through a serial chain, no probe required. It is the only way to test the connections hidden under a BGA (a ball-grid array: a chip whose connections are solder balls underneath it, where no probe can reach). Boundary scan alone covers 60–80% of device pins; combined with ICT it reaches over 99% structural fault coverage. The rules: the JTAG signals (TDI, TDO, TCK, TMS, and optional TRST) must be routed and accessible, the header at the board edge, and at least 80% of net connections between JTAG-capable pins for the scan to be effective. Analogue signals cannot be boundary-scanned, they still need probe access.

The economics, in one line. If your process makes 2% defective boards and your test covers 90% of defects, your customer defect rate is 0.2%, 2,000 bad boards per million. At 99% coverage, it is 0.02%, 200 per million. And the DFT failures themselves are the yield problem: one vendor-cited figure, attributed to IPC research, puts testability issues at 35–40% of PCB manufacturing yield losses, with boards designed without DFT needing 3–5× longer testing and showing 15–25% higher field failure rates.

DEMO

Example Labs is a made-up company we use for worked examples.

This board is made up. Call it the Example Labs sensor board: a 4-layer board, 100 mm by 60 mm, carrying the event camera from Article #2, the IMU from Article #4, and the microcontroller that runs the Kalman filter and the PID loop from Articles #3 and #4. Same schematic, same parts, two layouts. The targets are the article's own recommendations, not measured results.

Step 1, The hobby layout. The traces are 4 mil, under the 6–8 mil recommended width. The annular rings are 3 mil, under the 5 mil minimum, thinner than the drill's own wander. The parts are rotated every which way, packed tight, no fiducials, 90° corners. There are no test points at all: the designer planned to "test it by plugging it in." The board works on the bench. It works once. The factory's DFM check flags the traces and the rings; the assembly house flags the placement; the test house has nothing to probe. The fix is a respin, about a week to ten days for a simple board, plus the boards and parts bought twice.

Step 2, The production layout. The traces are 8 mil, the rings 6 mil, the corners 45°, the parts spaced 40 mil apart and all facing the same way, fiducials at the corners, traces 20 mil from the edge. Test points sit on the bottom side on a 100-mil grid, 1.0 mm pads, covering 96% of the nets, every power rail, every signal, every clock. The JTAG header sits at the board edge. The factory's DFM check passes. The assembly house places without a collision. The bed-of-nails fixture tests 200 nets in under a minute.

Step 3, What each one costs on a 100-board run. The factory's defect rate is 2%, two boards in a hundred are bad. The hobby layout is tested by eye and by plugging it in: about 50% coverage, so one bad board in a hundred escapes to the field. The production layout runs ICT at 97% coverage, so about 0.06 bad boards escape, one bad board in roughly 1,700. The hobby layout also pays the respin: a week to ten days of waiting, plus the re-fabrication. The production layout pays the fixture, but the fixture is a one-time cost, and it is the same fixture for every board after the first hundred.

Step 4, What each one costs to repair. The site's own rule is "made to be kept, instruments designed to be recalibrated, not replaced." The hobby layout's fault is under a BGA with no test points: the board is scrap. The production layout's fault is on a probed net: the test says which net, the technician reworks one part, and the board ships. The difference between a repairable board and a disposable one is the test points, and they cost nothing at design time, only the discipline to place them.

One circuit, two boards: same schematic and parts, different layout rules

Hobby layoutProduction layout
Traces and rings4 mil traces, 3 mil rings8 mil traces, 6 mil rings
PlacementParts rotated every way40 mil part spacing, aligned
Corners90°45°
Test accessNo test points, no fiducialsBottom-side test points on a 100 mil grid, fiducials, 96% of nets probed

The decision test

Four questions. They sort the boards that are ready for a production run from the boards that need a design review first:

  1. Does the board pass its own factory's DFM rules, trace width, spacing, annular ring, aspect ratio? No, the board is not ready, and the factory will tell you so at the worst possible time: after you have paid. Yes, the fabrication risk is handled.
  2. Can 95% of the nets be probed, and is there a test plan for the other 5%? No, the board's faults will reach the customer, and the first you hear of them is a return. Yes, the test risk is handled.
  3. Is there a second source for every part with a lead time over 12 weeks? No, a single component change after the freeze costs over $50,000 on average. Yes, the supply risk is handled.
  4. Can a technician open the product, probe the board, and replace one part? No, the board is disposable, and the site's own rule says instruments are made to be recalibrated, not replaced. Yes, the board is a product, not a consumable.

If the answers are yes, yes, yes, yes, the board is ready for a production run, and the review is a final check, not a rescue. If any answer is no, the job is a design review, and the honest thing is to do it before the boards are ordered, not after the first hundred arrive.

"But our factory checks the files", and two more fair doubts

"Our factory runs a DFM check, so we're covered." The factory's DFM check catches what the factory cannot build. It does not catch what the factory can build but cannot test, the nets with no test points, the BGA with no JTAG chain, the board with no fixture. And it does not catch the design decisions that cost you later: the part with one source and a 40-week lead time, the board that cannot be opened and repaired. The factory checks the board. The review checks the product.

"We only need ten boards." Then the rules still apply, and the cost of ignoring them is smaller, a week, not a month; a re-order, not a $46,000 respin. But the habits are the same, and the tenth board is where the habit pays: the test points that let you find the fault in minutes instead of hours, the fiducials that let the assembler place without a mistake. For ten boards, the review is an afternoon. For a thousand, it is the difference between shipping and not shipping.

"DFM slows us down." It slows the layout by a day and saves a respin of a week to ten days, or four weeks and $10,000–$50,000 on a complex board. The teams that simulate first need 3.0 spins to production; the teams that build and test need 4.6. The slowdown is a rounding error against the spins it prevents.

Sources

  1. Lifecycle Insights research, via Design007 and Tovan (2026)
    What we used

    Average 2.9 respins per project; 2.8 respins at $46,000 each (production-grade); build-and-test teams 4.6 spins to production vs simulate-first teams 3.0; simple prototype respin about a week to ten days; complex production respin about four weeks and $10,000–$50,000; budget two spins, treat a third as a process problem

  2. Accuris / Fuld & Company, Electronic Parts Intelligence Survey (N=439, March 2026)
    What we used

    68% make more than six component changes per design project; 40% experience over 20% design rework; 51% report over 11% of designs need component changes after the design freeze; 46% estimate the average post-freeze change exceeds $50,000 (up to $250,000 in aerospace, defence, automotive); engineering rework $15,000–$80,000 per change event; total per forced redesign $135,000–$930,000+; semiconductor lead times nearly doubled February–March 2026, top-component lead times reaching 40 weeks

  3. HiElectronic, "The True Blind Buried Via PCB Cost Breakdown"
    What we used

    HDI respin $3,500–$9,000 direct and schedule cost; NRE $600–$1,800 per design, charged again per revision; factory yield loss 3–8% scrap absorbed internally; tooling retained 12–24 months

  4. YourPCB, "PCB DFM Design Rules" (IPC-2221 / IPC-6012)
    What we used

    Minimum trace width 5 mil Class 2 / 4 mil Class 3, recommended 6–8 mil; minimum spacing 5 / 4 mil, recommended 6–8 mil; trace-to-edge 10 mil minimum, 15–20 mil recommended; via drill 8 / 6 mil, recommended 10–12 mil; annular ring 5 mil minimum; via aspect ratio 8:1 / 10:1, recommended 6:1; solder mask clearance 3 mil; copper-to-edge 10 mil minimum, 15 mil recommended

  5. MorePCB, "DFM Guidelines Every PCB Engineer Should Know"
    What we used

    40 mil minimum component spacing for pick-and-place; 100 mil for tall parts; 200 mil for heat-generating parts; 100 mil around mounting holes; 20 mil trace-to-edge; annular ring 5 mil minimum, 7–10 mil preferred; 8:1 maximum aspect ratio; voltage clearances 0.1 mm (0–30 V), 0.4 mm (150–300 V), 1.5 mm (400–500 V); 90° corners cause acid traps

  6. Schemalyzer, "PCB Design Rules Every Engineer Should Know" (2025)
    What we used

    IPC-2221 Table 6-1 clearances (0–15 V: 0.05 mm internal, 0.1 mm external uncoated; 301–500 V: 0.25 mm internal, 2.5 mm external); power traces 1 A = 20 mil, 3 A = 60 mil (1 oz copper, 10 °C rise); drill registration tolerance ±0.075 mm (3 mil); do not place test points under ICs

  7. AtlasPCB, "PCB Design for Testability (DFT)" (IPC-9252A, IEEE 1149.1)
    What we used

    Test point pad minimum 0.9 mm (35 mil), recommended 1.0 mm (40 mil); pitch minimum 1.27 mm (50 mil), recommended 2.54 mm (100 mil); bottom-side priority; net coverage target ≥95% (≥98% Class 3, ≥90% commercial); coverage vs escape: visual only 40–60% / 500+ ppm, flying probe 85–92% / 50–100 ppm, ICT 95–98% / 10–30 ppm, ICT + boundary scan 98–99.5% / 2–10 ppm, plus functional test 99.5–99.9% / under 5 ppm; boundary scan target ≥80% of nets between JTAG-capable pins

  8. AndwinPCB, "Design for Test (DFT)"
    What we used

    DFT issues cause 35–40% of PCB manufacturing yield losses (vendor-cited, attributed to IPC research); boards without DFT need 3–5× longer testing and 15–25% higher field failure rates; ICT 200–400 nets in 15–90 s; flying probe 3–10 min; test point minimum 40 mil (1.0 mm) per IPC-2615

  9. Industrial PCB Manufacturing, "Design for Testability"
    What we used

    2% defective boards at 90% coverage gives 0.2% customer defect rate; 99% coverage gives 0.02% (200 per million); bed-of-nails fixture $10,000–$30,000 per board revision

More in design notes

The Profile That Doesn't TombstoneYour boards come back with tombstoned parts, cold joints, and voids, and the assembler says the profile is fine, because the oven's setpoints are fine and the board's temperatures are not. Here is the reflow profile as four zones, measured on the board, with the numbers.17 minAluminium, Steel, or PlasticAluminium, steel, or plastic? The material decides the cost, the weight, the heat, the EMI, and whether the product survives outdoors, and the wrong one fails in under two years. Three axes, four plastics, three metals, and the arithmetic for Example Labs' field sensor hub.20 minCNC Machining DFMA 30 mm deep pocket with R2 corners forces a 4 mm end mill that deflects 16 times more than the tool the job needs, and a drawing with ±0.01 mm on every dimension costs eight times the same part at ±0.5 mm. Eight design rules, each with a number, and a machined sensor hub worked from $1,370 a part to $430.25 minEnclosures That Survive the Real WorldIP67 is a laboratory test on a new unit, for thirty minutes. It says nothing about year two, when the gasket has taken a compression set and the box has baked in the sun. A sealed 100 × 60 × 40 mm box with 5 W inside runs 25 °C hotter than the air around it. The enclosure is not a box, it is a promise: keep the water out, let the heat out, and open again when something breaks.14 minInjection Moulding DFMInjection molding design determines 70–80% of final part cost and quality before production begins. The five rules are simple: a uniform 1.5–3.0 mm wall, 1–3° of draft per side, ribs at half the wall, bosses at six-tenths, and shrinkage built into the cavity, PP shrinks 1.0–2.5%, so a 100 mm part needs a cavity of 101–102.5 mm. Roughly 70% of first-submission designs contain at least one feature that would cause a defect or force an expensive mould action, and one post-tooling modification typically costs $2,000–$20,000 and 2–6 weeks. The fix is a CAD edit, not a tool edit.19 min