Design for Test
Design for Test: The Fixture That Catches the Fault Before the Customer Does
A board that passes the bench and fails in the field is usually untestable by design, no test points, no JTAG chain, no fixture. The IPC-9252A rules, the four test methods, the coverage union, and what a fixture is worth on a real board.
Picture a production run of 500 industrial controllers. Every board passes visual inspection, AOI, and X-ray. Then firmware touches an untested I/O rail, and the board fails.
Or the opposite: a board passes a 90-second functional test, ships, and later fails in the field, because one wrong 10 kΩ resistor was never isolated at component level, and the resistor looks perfect under a camera. A 1 kΩ resistor placed where a 10 kΩ belongs has the same package, the same solder joints, and the same silhouette. The camera cannot tell them apart.
The board can.
That is the gap design for test exists to close. Design for test, DFT, is the practice of building testability into the board from the start: accessible test points, a JTAG chain, power-rail isolation, and the hooks the firmware needs to test itself. It costs almost nothing in board area, typically 2–4% overhead, and it is the difference between a fault that is found in seconds on the production line and a fault that is found by the customer.
This article is about the board you can actually test: the rules for the test points, the four test methods and what each one owns, the coverage union that decides what escapes, and the fixture that catches the fault before the customer does.
What design for test is, in plain language
A board has two kinds of correctness. The first is whether the design is right, whether the circuit does what it is supposed to do. That is verification, and it happens once, before anything is built. The second is whether each physical copy came out right, whether this board, soldered by this line, has the right parts in the right places with sound joints. That is manufacturing test, and it happens to every board.
DFT is the set of features that makes the second question answerable. The problem it solves is access: a modern board hides its connections under BGA packages, behind 0201 passives, and inside multilayer stackups, and a tester that cannot touch a net cannot measure it. DFT gives the tester something to touch.
The three access features are physical test points (exposed copper pads on the nets), boundary scan (test logic already built into the digital chips, reached through the JTAG port), and built-in self-test (firmware that exercises the board and reports what it finds). The fixture, the bed of nails, is the machine that uses the first of these, and the article that follows is about all three, in the order a production line actually uses them.
The cost of not doing it is not theoretical. Boards without proper DFT planning need 3–5× longer testing time and show 15–25% higher field failure rates than designs with structured test access (Andwin, citing IPC research), and Andwin, a PCB manufacturer, so a vendor-cited figure, flagged as such, puts DFT issues at 35–40% of PCB manufacturing yield losses. A high-frequency, high-speed board in Embedded's account had no room for test probes on its 0201 passives and ended up with 30% test coverage, where a typical PCB should be 70–80% accessible with testing probes.
The board went through manufacturing missing the critical tests it needed, and the fix, leaving a strip of solder mask off the end of the component pad so a probe could hit exposed copper, had to be designed in after the fact, at the CAD level, with several steps. DFT is a schematic and layout decision. A production fixture cannot recover observability the PCB never exposes.
The rules: test points per IPC-9252A
The rules come from IPC-9252A, the standard for electrical testing of bare boards, as the guides apply it to assembled boards. They are small numbers, and they are worth memorising, because they are the numbers the fixture is built to.
Pad diameter: 0.9 mm minimum, 1.0 mm recommended. A 0.9 mm pad is 35 mil; 1.0 mm is 40 mil. The pad is the target the spring-loaded probe has to land on, and a smaller pad means a smaller probe, a weaker spring, and a higher chance of a miss. At a 2.54 mm pitch, a 1.00 mm pad gives a near-zero contact-miss rate over high-volume runs; shrink the pitch to 1.27 mm and the pad has to shrink to 0.60 mm, and the probe-hit probability falls to 98.5%. The recommended size is the one that makes the miss rate negligible.
Centre-to-centre pitch: 1.27 mm minimum, 2.54 mm recommended. 1.27 mm is 50 mil; 2.54 mm is 100 mil. The pitch is the distance between two test points, and it decides the probe that fits: a 2.54 mm grid takes heavy-duty probes with 1.37 mm barrels and springs up to 2.8 N; a 1.27 mm pitch restricts the barrel to 0.68 mm and cuts the spring force to 1.1 N. The recommended grid is the one that takes the cheap, durable probe.
Net accessibility: aim for ≥95%. Every signal net, power rail, and ground connection should have at least one test point. The 5% allowance covers the nets that physically cannot route to an accessible pad, BGA inner balls on via-in-pad, for example. For IPC Class 3 (high-reliability) applications the target is ≥98%; for commercial products, ≥90% is acceptable. Test access percentage is accessible nets divided by total nets, times 100, and it is the number the whole article is about, because every coverage figure that follows is a fraction of it.
Placement. Test points go on the bottom (secondary) side for a bed-of-nails ICT fixture, so the board drops onto the probes with the components facing up. They align to a 2.54 mm grid when possible, grid-aligned points reduce fixture cost by $200–$800 per design, because the fixture drill plate uses standard grid patterns. Clearances: 1.0 mm minimum (1.27 mm recommended) from components, 2.5 mm minimum (3.0 mm recommended) from the board edge, and no solder mask over the pads. One-side access, where possible, halves the fixture cost: a double-sided fixture needs probes on both sides and a board that can be pressed from both sides.
Example Labs is a made-up company we use for worked examples.
Probe force: 100–200 g per probe. Each spring-loaded probe pushes with 100–200 grams, so a board with 1,000 test points experiences 100–200 kg of total force. The board and its support structure have to handle that without deflecting more than 2 mm, which is why the fixture has support pins under the board and why the probe count is a mechanical budget, not just an electrical one. The demo below works the force budget for the Example Labs board.
The four methods, and what each one owns
There are four test methods, and the useful concept is failure ownership: give each method the defects it can isolate quickly, and give the behaviours that only appear when the circuit is powered to the functional test. The methods split into two families. Inspection (AOI and X-ray) checks that parts are present, aligned, and soldered correctly, it sees geometry, not function.
Electrical test (flying probe and ICT) applies signals and measures the response, so it catches a wrong-value resistor or an open that looks perfect under a camera. And the study the guides keep citing is the IPC conference paper from 2002: of the defects found by circuit test on four products on a high-volume line, 58–61% could not be detected by AOI, and without circuit test, those defects would have reached the customer.
In-circuit test (ICT), the bed of nails. A custom fixture presses hundreds of spring-loaded probes onto the test points at once, then measures every component in parallel: resistance, capacitance, continuity, and selected powered behaviour. It tests 200–400 nets at once, in 15–90 seconds per board, and on a well-designed board it reaches 85–95% net coverage, the top end, 95–98%, is the full-access figure the Atlas guide quotes for a board with complete test-point access. The fixture is the catch: a one-time investment of $3,000–$15,000 for a typical board (Andwin), and the range runs to $30,000 a revision for complex boards, the #5 file recorded $10,000–$30,000 per revision from an earlier source, and JLCPCB puts complex fixtures at $10,000–$50,000. Change the board and you often rebuild the fixture, which is why ICT pays off at volume and not for one-off prototypes: at 10,000 boards the per-board test cost falls to about $0.10, against $0.50–$2.00 for flying probe. ICT is the high-volume workhorse, and it is the method that gives the strongest fault isolation, a failure points to a net or a component, often down to a specific pin.
Flying probe, the fixtureless electrical test. Two to eight motorised probes move across the board, touching test pads and pins one at a time, guided by the netlist. There is no fixture, so a board revision means reloading the program instead of scrapping tooling, and the probes can reach both sides of the board and microvia land pads a pogo pin cannot. The trade is speed: probing every net sequentially takes 3–10 minutes per board, so it does not scale to thousands of units, and coverage is 70–85% of nets. It is the right method for prototypes, for production under 500 units a month, and for boards whose design changes frequently. The break-even against ICT is usually 300–1,000 boards: below that, the fixture is not justified; above it, the fixture's seconds-per-board win.
Boundary scan (IEEE 1149.1 JTAG), the test logic already in the chips. JTAG-compliant devices contain a boundary scan register on every I/O pin, a shift-register cell between the physical pin and the chip's core logic, controlled through a Test Access Port of four required signals (TDI, TDO, TCK, TMS) plus an optional TRST. The devices are wired in a daisy chain, TDO of one to TDI of the next, and the test drives a pattern out of one device's pins and captures it at another's, testing the interconnect without any physical probe. That is what makes it the route for BGA-dense boards: the solder joints are hidden under the component body, and no probe can reach them, but the boundary scan cells can. Coverage is 60–80% of device pins, the cost is low (software), and the speed is medium (30–120 seconds). It complements rather than replaces physical test points, it cannot test an analog net or a power rail, because those have no boundary cell, and it is also the field-diagnostics route: the same JTAG port that tests the board on the line can test it in a customer's rack, years later.
Functional test, the final gate. The board is powered and exercised as a working system: voltage sequencing, firmware boot, communication buses, sensors, calibration, current draw. It is the only method that proves the product works, and it is the only method with weak fault isolation: a board that fails functional test has failed somewhere, and the technician spends minutes or hours finding where. Functional test proves a unit works; it does not identify whether the cause is an open, a short, a wrong component, or a programming error. That is why production lines pair it with ICT as the final gate, ICT confirms the board was built correctly, functional test confirms it works, and why the common sequence runs AOI after reflow, then X-ray for hidden BGA joints, then ICT or flying probe for electrical verification, then functional test as the final gate.
Built-in self-test, the fifth method, inside the firmware. The RTOS piece built the hooks: each critical task sets a flag per successful cycle, and one supervisor task checks every flag before it kicks the watchdog. The same hooks serve the fixture: the board powers up, runs its self-test, and reports the result over the debug port. BIST is the cheapest coverage there is, it costs firmware, not fixture, and it is the only method that works in the field, where there is no fixture at all. The fixture's job, in the demo, is to program the firmware and run the self-test; the secure-boot piece built the path, and the chain of trust means the fixture must present a signed image, or the board will not boot. The fixture is part of the trust chain, not outside it.
Coverage is a union, not a sum
The most common coverage error is adding the methods' percentages together. Coverage is a union: the same fault detected by three methods counts once. A fault is covered if any method catches it, and the coverage of the stack is the size of the union of the methods' covered sets, not the sum of their percentages.
The defect-escape table, from the Atlas guide, shows what the union buys. Visual inspection only: 40–60% coverage, 500+ ppm escape. AOI plus visual: 70–80%, 200–300 ppm.
Flying probe (nets only): 85–92%, 50–100 ppm. ICT (full access): 95–98%, 10–30 ppm. ICT plus boundary scan: 98–99.5%, 2–10 ppm.
ICT plus boundary scan plus functional test: 99.5–99.9%, under 5 ppm. Each step up the table is a method added to the union, and each step cuts the escape rate by roughly an order of magnitude, which is the same law the chip industry writes as the Williams relation: defect level = 1 − yield^(1 − coverage), so at 90% yield, 90% coverage ships about 10,000 bad chips per million, 99% coverage about 1,000, and 99.9% about 100. The Williams formula is the chip-level version of the same law, and the article cites it as such; the board-level numbers are the Atlas table's.
The point is the same in both: each extra nine of coverage cuts the shipped defects roughly tenfold, which is why test teams fight over fractions of a percent, and why the demo's 97.5% is worth more than it looks.
The other number that decides the economics is the cost of the escape. A defective board that escapes to the field costs 10–100× more to fix than one caught on the line (Atlas), the repair, the shipping, the support time, and the customer. The demo prices it at the conservative end, 10×, and the fixture still pays for itself in the first run.
The build: the fixture, and what it costs
The fixture is a mechanical platform, and it is built for one board. A frame holds the board; a vacuum or a mechanical press pulls it down onto a grid of spring-loaded probes, the bed of nails, arranged to touch every test point at once; a press plate holds it flat against the probe force. The probes are pogo pins, beryllium copper with gold plating, typically lasting 100,000–500,000 test cycles.
The fixture is custom-drilled for one board design, which is why a board revision often means a rebuilt fixture, and why the design has to be stable before the fixture is ordered.
The cost has two buckets: the one-time fixture and program, and the recurring per-board test cost. The fixture is the big number, $3,000–$15,000 for a typical board, up to $30,000 for a complex one, and it is why ICT pays off at volume and not for prototypes. The per-board cost collapses with volume: about $0.10 a board at 10,000 units, against $0.50–$2.00 for flying probe.
The lead time is 2–6 weeks for the fixture, against hours for a flying-probe program. The decision rule the guides give is simple: choose ICT when the volume exceeds 1,000 units a month, the design is stable with minimal engineering changes, and the test time has to be under 90 seconds; choose flying probe when the volume is under 500 units a month, the design changes frequently, or the board has insufficient test access for a fixture.
The fixture also has to be designed for the board's mechanics, not just its electrics. The probe force budget is 100–200 g per probe, and the board must not deflect more than 2 mm under the total force, so the fixture has support pins under the board, and the probe count is a mechanical budget. A board with 1,000 test points takes 100–200 kg of force, and a thin board without support pins bows, the probes lose contact, and the test fails on boards that are fine.
The demo works the force budget for the Example Labs board.
DEMO
Example Labs' field sensor hub is a 100 × 60 mm, 4-layer board with 120 nets, built around the STM32H7 from the cluster (512 kB SRAM, 2 MB flash, 48 MHz clock, BLE radio at 2.4 GHz, firmware image 384 KiB). The fault model is 120 modelled faults, one per net, a stated assumption, matching the pipeline's worked example, and the arithmetic was run in Python before writing.
Step 1, the access. As built, 114 of the 120 nets have a test point: 95.0% net accessibility, meeting the ≥95% target. The 6 inaccessible nets are BGA inner balls and via-in-pad under the STM32H7, the 5% the guides allow for. The test points are 1.0 mm pads on a 2.54 mm grid, on the bottom side, with 1.27 mm clearance from components and 3.0 mm from the board edge, the recommended column of the IPC-9252A table, not the minimum. The board area the test points cost is under 4%, the guides' typical overhead.
Step 2, the coverage union. Each method covers a fraction of the 120 modelled faults, at the midpoints of its verified range: ICT 90% (85–95%) → 108 faults; flying probe 80% (70–85%) → 96 faults; boundary scan 70% (60–80% of device pins) → 84 faults. The union is 117 of 120, or 97.5%: ICT catches 108, the flying probe adds 6 the ICT fixture cannot reach (both sides, and vias under components), and boundary scan adds 3 (the BGA inner-ball interconnects). The 3 faults that escape are analog and power nets under the BGA, no boundary cell, no test point, no method that can reach them. They are the cost of the 5% the board could not expose, and the only way to catch them is the functional test, which can see the effect of a bad power net even when it cannot probe the net itself.
Step 3, what escapes, and what it costs. The line builds 10,000 boards in a run, and the first-pass defect rate is 2%, a stated assumption, so 200 boards in the run carry a manufacturing fault. At the full stack's 97.5% coverage, 5 faulty boards escape; at ICT alone (90%), 20 escape; at flying probe alone (80%), 40 escape; at visual-only (50%, the midpoint of 40–60%), 100 escape. The cost: in-factory rework is $50 a board (stated), and a field return costs 10×, $500, the conservative end of the guides' 10–100× range. So the escapes cost $2,500 at the full stack, $10,000 at ICT alone, $20,000 at flying probe alone, and $50,000 at visual-only. The fixture and program cost $12,000, an $8,000 fixture, inside Andwin's $3,000–$15,000 range, plus $4,000 of programming, both stated assumptions, and the stack against visual-only saves $47,500 a run. The fixture pays for itself in the first 10,000-board run. Against ICT alone the stack saves $7,500 a run, and boundary scan is software cost only, so the incremental cost of the stack over ICT alone is small and the saving repeats every run.
Step 4, the time. The electrical step, per board: ICT 30 seconds (inside the 15–90 second range) against flying probe 5 minutes (the midpoint of 3–10). Over a 10,000-board run that is 83 hours of test-station time against 833 hours, a 750-hour difference, which is the throughput argument for the fixture. The full stack per board is 30 seconds of ICT, 60 seconds of boundary scan (inside 30–120), and 3 minutes of functional test (the Andwin automotive example's figure), 4.5 minutes in total. The functional test dominates the stack's time, and it is the same regardless of the electrical method, so the choice between ICT and flying probe is the choice between 30 seconds and 5 minutes for the electrical step. The Andwin automotive example makes the same point at a different scale: flying probe on 120 analog and power nets (40% of nets) in 8 minutes, boundary scan on 280 digital interconnects (55% of nets) in 90 seconds, functional test in 3 minutes, 12 minutes a board, 96% fault coverage, and the coverage is the union, not the sum.
Step 5, the force budget. The fixture has 114 probes, one per test point, at 150 g each, the midpoint of the 100–200 g range, for a total of 17.1 kg of force pressing the board down. The board is 1.6 mm thick (stated), and the fixture's support pins hold the deflection under the 2 mm the guides allow. A board with 1,000 test points would take 100–200 kg, and a thin board without support pins would bow and fail the test on boards that are fine. The force budget is a line in the fixture design, next to the test-point count.
The traps
Designing test points last. The test points are placed after the layout is done, and there is no room left, so the board ships with 30% coverage where a typical board should be 70–80% accessible, and the fix is a solder-mask strip over a via pad, designed in after the fact. Test points are placed in the schematic and in placement, not in the layout's leftover space. The rule is that DFT begins in the schematic: the nets that need test points are named before the board is routed.
Test points on the wrong side. The points are on the component side, and the bed-of-nails fixture needs them on the bottom. The fixture has to be double-sided, or the board has to be flipped, and either way the cost doubles. One-side access, on the bottom, is the default, and the decision belongs in the layout review, not in the fixture quote.
No test points on power rails. The signal nets are covered and the power rails are not, so a shorted rail or a wrong-value bulk capacitor is invisible to the fixture. Every power rail and every ground connection needs at least one test point, the ≥95% target counts them, and the power-rail isolation the guides recommend (a jumper or a 0 Ω link that lets the fixture power a rail on its own) is what lets the ICT measure a rail without the rest of the board loading it.
Trusting functional test alone. The board passes functional test and ships, and a wrong 10 kΩ resistor fails in the field, because functional test proves the unit works and does not identify whether the cause is an open, a short, a wrong component, or a programming error. The functional test is the final gate, not the only gate, and a failure at the final gate costs minutes or hours of diagnosis per board, against seconds at the ICT station.
Over-testing. The stack is ICT plus boundary scan plus functional test plus X-ray plus AOI on a 500-unit run, and the test cost eats the margin. The stack is chosen by volume and risk, not by habit: prototypes get AOI plus flying probe; stable production above roughly 1,000 units gets the ICT fixture; BGA-dense boards get boundary scan; every board gets functional test as the final gate. No single method achieves 100% fault coverage economically, and the optimal strategy combines them, but only the ones the volume and the risk justify.
The decision test: five questions
A team choosing a test strategy can answer five questions in an afternoon. If it cannot answer three of them, it has not chosen a strategy, it has chosen a fixture quote.
- What is the net accessibility, accessible nets divided by total nets, and is it above 95%? If the number is below 90%, the board is not testable, and the fix is a layout change, not a fixture.
- Which nets have no test point, and why? If the answer is "the BGA inner balls", boundary scan is the route. If the answer is "we ran out of room", the layout is the problem.
- What is the volume, and how often does the design change? Above 1,000 units a month with a stable design, the ICT fixture pays for itself. Below 500 units a month, or with frequent changes, flying probe is the answer.
- What does each method own? If the answer is "functional test catches everything", the team has confused proving the unit works with finding the fault, and the first field failure will teach the difference.
- What is the probe force budget, and does the fixture support the board? If the board deflects more than 2 mm under the probes, the test fails on boards that are fine, and the fixture, not the board, is the fault.
Objections
"Our boards are simple, we don't need a fixture." A simple board still has wrong-value resistors, and a 1 kΩ resistor where a 10 kΩ belongs looks perfect under a camera. The question is not whether the board is simple. It is whether the volume justifies the fixture, and at 500 units, a flying probe and a functional test are the answer, not nothing.
"Test points cost board area we don't have." They cost 2–4% of the board area, typically, and the cost of not having them is 3–5× longer testing time and 15–25% higher field failure rates. The area is the cheapest insurance on the board. The expensive version is the respin that adds the test points after the first production run fails.
"Boundary scan replaces the fixture." It complements the fixture; it does not replace it. Boundary scan covers 60–80% of device pins, and it cannot test an analog net or a power rail, because those have no boundary cell. The highest fault coverage comes from combining physical probe access with boundary scan, over 99% structural fault coverage, against 95–98% for ICT alone.
"AOI already catches the defects." AOI catches placement and solder defects. It cannot catch a wrong-value resistor, an open under a BGA, or a short between two nets that look identical from above. The IPC conference paper's number is the answer: 58–61% of the defects circuit test found could not be detected by AOI, and without circuit test those defects would have reached the customer.
"We'll just test in the field." The field is the most expensive test station there is: 10–100× the cost of catching the fault on the line, plus the customer. A board that is tested in the field is a board whose test strategy is the customer's patience.
Sources
- Atlas PCB, "PCB Design for Testability (DFT)"
What we used
IPC-9252A test-point table: pad diameter 0.9 mm minimum / 1.0 mm recommended; centre-to-centre pitch 1.27 mm minimum / 2.54 mm recommended; clearance from components 1.0 mm / 1.27 mm; clearance from board edge 2.5 mm / 3.0 mm; pad height flat / <0.5 mm protrusion; net coverage target ≥95%; grid alignment to 2.54 mm reduces fixture cost by $200–800 per design; probe force 100–200 g per probe (1,000 test points = 100–200 kg total, deflection <2 mm); defect-escape hierarchy, visual-only 40–60% / 500+ ppm, AOI + visual 70–80% / 200–300 ppm, flying probe 85–92% / 50–100 ppm, ICT full access 95–98% / 10–30 ppm, ICT + boundary scan 98–99.5% / 2–10 ppm, ICT + boundary scan + FVT 99.5–99.9% / <5 ppm; Class 3 target ≥98% test access, commercial ≥90% acceptable; DFT costs typically 2–4% board area; field fixes cost 10–100× more
- Andwin, "Design for Test (DFT): Test Points & Boundary Scan in PCB"
What we used
DFT methods comparison: ICT 85–95% (nets), 15–90 s, fixture $3,000–$15,000, volume >1,000/month; flying probe 70–85% (nets), 3–10 min, no fixture, <500 units/month; boundary scan 60–80% (device pins), low cost (software), 30–120 s, BGA-dense, field diagnostics; functional test 100% (application), high cost; DFT issues cause 35–40% of PCB manufacturing yield losses (vendor-cited); boards without proper DFT need 3–5× longer testing time and show 15–25% higher field failure rates (Andwin, citing IPC research); ICT tests 200–400 nets at once; ICT delivers 90–95% net coverage; boundary scan achieves 60–80% on boards with sufficient JTAG density (≥60% of digital ICs supporting boundary scan → 70–80% structural coverage); worked automotive example, flying probe 120 analog/power nets (40% of nets) in 8 min, boundary scan 280 digital interconnects (55%) in 90 s, functional test 3 min, total 12 min/board, 96% fault coverage; no single method achieves 100% fault coverage economically
- Andwin, "PCB Testing Methods: AOI, X-Ray, Flying Probe & ICT"
What we used
ICT on a well-designed board reaches 85–98% of nets and component values; fixture one-time NRE roughly $800 to several thousand for a simple board, far more for complex ones; flying probe 3–8 min per board; ICT 10–60 s; flying probe proto to ~1,000 units, ICT 1,000+; ICT measures a resistor's value but never proves the board boots
- JLCPCB, "Bed of Nails Testing for High-Volume PCBs"
What we used
ICT defect detection over 95%, some systems 98% or more; flying probe 1–5 min per board vs bed-of-nails under 10 s; fixture $10,000–$50,000 depending on complexity and test-point count; per-board test cost ~$0.10 ICT vs $0.50–$2.00 flying probe; defect coverage 90–98% ICT vs 70–80% flying probe; setup lead time 2–6 weeks vs hours; pogo pins last 100,000–500,000 test cycles
- HI Electronic, "Electrical Testing PCB"
What we used
And "ICT (In-Circuit Test)", flying probe vs ICT fixture cost $0–$300 programming vs $100+ simple / $1,000–$3,000+ complex; cost per board $5–$25 flying probe vs $0.50–$4.00 ICT after amortisation; break-even 300–1,000 boards; minimum test pad 0.30 mm flying probe vs 0.80–1.00 mm ICT; BGA-heavy boards: accessible coverage 60–80%; ICT cannot confirm the product functions, power sequencing, firmware behaviour, RF performance lie outside its scope; production sequence AOI → X-ray → ICT/flying probe → FCT
- Keysight, "In-Circuit Test Systems"
What we used
ICT detects over 90% of manufacturing defects; test cycle often under a minute; ideal for medium- to high-volume boards with moderate to high component counts
- Atlas PCB, "Boundary Scan (JTAG) Testing for PCB Assembly"
What we used
The access problem: BGA joints hidden under the component body, no physical probe can reach them; IEEE 1149.1 defines the boundary scan register, the TAP, the TAP controller, and the instruction register; TAP signals TDI/TDO/TCK/TMS plus optional TRST; daisy-chain topology; infrastructure test, interconnect test (opens, shorts), and the EXTEST instruction
- YourPCB, "ICT vs Functional Test in PCB Assembly"
What we used
ICT finds assembly structure (opens, shorts, wrong values, polarity, missing parts); functional test proves behaviour (firmware boot, I/O, calibration, comms, sensors) with weaker fault isolation; the failure-ownership concept; a board can pass a 90-second functional test and ship, then fail because one wrong 10 kΩ resistor was never isolated at component level
- Jiahon, "PCBA Test Strategy"
What we used
IPC conference paper (2002): 58–61% of the defects found by circuit test on four high-volume products could not be detected by AOI, and without circuit test those defects would have reached the customer
- Embedded.com, "Overlooking design-for-test can lead to costly PCB design rework"
What we used
Without DFT, coverage can fall to 30% where a typical PCB should be 70–80% accessible; the 0201 passive example (0.325 mm pad, no room for a probe) and the solder-mask-strip fix; DFT's main objective is contact points for the probes
- Chipfield Guide, "Design for test"
What we used
The Williams relation: defect level DL = 1 − Y^(1−T); at 90% yield, 90% coverage ships ~10,000 bad chips per million, 99% coverage ~1,000, 99.9% ~100; each extra nine of coverage cuts the bad chips tenfold. Chip-level formula, cited as such
- Example Labs figures from the cluster: the field sensor hub (100 × 60 × 40 mm
What we used
#7); the STM32H7 (512 kB SRAM / 2 MB flash, 48 MHz clock, #9); the BLE radio at 2.4 GHz (#12); the firmware image 384 KiB (#28); the 48 V bench pack (13S1P NMC, 3.5 Ah/cell, 48.1 V nominal, 168 Wh, #20/#23)