TDEB Top Bar

Trending
Semiconductor Industry News — The Digital Electronics Blog
Follow Us

TDEB Masthead

The Semiconductor Hub

Semiconductors.  Chips and Systems.  Innovation.

Indexed by leading Search Engines, AI and referenced by Engineers worldwide 500+ Articles 10M+ Pageviews 30+ Reports 50+ Citations

AEC-Q100 Test by Test: What a Pass Proves and What It Leaves Open

Murugavel Ganesan
by
0
Share Facebook X LinkedIn WhatsApp Email
Automotive chips on a test board inside a temperature chamber during reliability stress testing

Ask a supplier whether a chip is automotive grade and the answer is usually "AEC-Q100 qualified". That phrase stands for a set of stress tests the supplier runs itself, with no certification body behind it. The main stresses use 77 parts from each of 3 lots, 231 in all, with zero failures allowed, and the standard says plainly that this sample is meant to expose design, construction and material problems, not low parts-per-million escapes.

This post walks through the current document, Rev-J1 of 24 March 2026, one test group at a time. For each group it covers what the stress does, which field failures it points at, and what it leaves open. An illustrative HTOL calculation then shows why the junction temperature in the oven can matter more than the grade on the datasheet.

Key takeaways

  • AEC-Q100 is a supplier-run qualification. AEC states that it runs no certification board, and the user is responsible for checking the data.
  • Grades 0 to 3 set the ambient range, and the hot end of each grade sets the HTOL temperature.
  • Zero failures in 231 parts is about 1% at 90% confidence, by AEC's own description. Low-rate defects are left to process control and screens.
  • A standard 1,000-hour HTOL does not cover every mission profile. In AEC's own example, a 125 °C junction would need about 1,390 hours.
  • Wear-out data (Group D) carries no sample size or pass count in the standard. Ask for it.

What the label does and does not mean

Three clauses are worth reading before any test table. Section 1.3.2 says there are no AEC "certifications". Each supplier qualifies its own part, weighs customer requirements and hands the data to the user, who has to verify it. So "AEC certified" on a sales sheet is loose wording. Our post on designing hardware that survives real-world failures lists that misreading among the common mistakes.

Section 2.2 sets the order of precedence: purchase order, then the agreed device specification, then AEC-Q100, then its reference documents, and the supplier's datasheet last. A purchase order can add to the standard, but a part cannot be called qualified if the order waives or weakens its requirements.

Section 3.3 sets the pass rule. Every failure needs a root cause, and the part counts as qualified only once corrective action has been proven effective and an 8D report is complete. Failures caused by mishandling, such as ESD damage during testing, may be discounted, though they must still be reported.

Temperature grades

All four grades start at -40 °C. Only the hot end differs. The table below sets each grade beside the two stresses that scale most visibly with it.

GradeAmbient hot endHTOL for 1,000 hoursTemperature cycling
0+150 °C+150 °C ambient-55 °C to +150 °C, 1,500 cycles or equivalent
1+125 °C+125 °C ambient-55 °C to +150 °C, 1,000 cycles or equivalent
2+105 °C+105 °C ambient-55 °C to +125 °C, 1,000 cycles or equivalent
3+85 °C+85 °C ambient-55 °C to +125 °C, 500 cycles or equivalent

Grade 0 cycling is one of the changes older data can miss: Rev-H (2014) asked for 2,000 cycles and the current text asks for 1,500. Check which revision a qualification report was written against. Rev-J is dated 11 August 2023, and the J1 update changed the Appendix 2 certificate of design and construction.

The tests, group by group

Not every test applies to every part, and the notes in the standard's main table decide by package and device type. The links to field failures here are our own reading, using the models and checks the standard names.

Group A: accelerated environment

These stresses attack the package and its interfaces with moisture and temperature swings. Most use 77 parts from 3 lots, with zero failures allowed.

  • Preconditioning (A1). Surface mount parts go through a moisture soak and reflow before the other package stresses, to mimic board assembly. Moisture level 3 is the minimum accepted. Nine devices are then checked for delamination with acoustic microscopy.
  • THB or biased HAST (A2). Humidity under bias: 85 °C and 85% humidity for 1,000 hours, or HAST at 130 °C for 96 hours or 110 °C for 264 hours. AEC's example model for this test is bond pad corrosion.
  • Autoclave, unbiased HAST or temperature-humidity (A3). Moisture without bias. The current text strongly prefers unbiased HAST to autoclave, and an appendix note warns that autoclave tends to reveal mechanisms not seen in normal use.
  • Temperature cycling (A4). Conditions are in the grade table. Afterwards, five decapped devices from one lot get wire bond pull tests, and nine devices get a delamination check. AEC's example links 54,750 engine on/off cycles over 15 years, with a 76 °C swing, to 1,034 test cycles using a Coffin-Manson exponent of 4.
  • Power temperature cycling (A5). This is only for parts with repeated power changes of at least 1 W, rise times under 0.1 s and junction swings of at least 40 °C, such as smart power switches. It uses 45 parts from 1 lot, and thermal shutdown must not occur.
  • High temperature storage (A6). An unpowered bake on 45 parts from 1 lot, at 150 °C for 1,000 hours or 175 °C for 500 hours for Grade 1 plastic parts.

Group B: accelerated lifetime

  • HTOL (B1). The part runs biased at the maximum supply voltage for its guaranteed parametrics, with 77 parts from 3 lots. It is tested before and after at room, cold and hot. Rev-J added advice to analyse drift in key parameters afterwards, where applicable, to confirm guard bands.
  • ELFR (B2). This looks for early-life failures on 800 parts from each of 3 lots. Under method Q100-008, Grade 1 runs 48 hours at 125 °C or 24 hours at 150 °C. Generic data is allowed, and for a part qualified for a single user, that user may accept part average testing in its place.
  • NVM endurance, retention and operational life (B3). This applies only to devices with non-volatile memory. Such parts must also receive endurance preconditioning before HTOL.

Group C: package assembly

These are mostly measurements rather than pass or fail stresses, and many are judged on capability, with Cpk above 1.67. Wire bond shear and pull (C1, C2), solderability (C3), physical dimensions (C4), solder ball shear (C5) and lead integrity for through-hole parts (C6) are the long-standing set. Bump shear (C7) arrived with Rev-J's new flip-chip BGA coverage, using 20 bumps or pillars from at least 5 devices.

Group D: die fabrication wear-out

Electromigration (D1), dielectric breakdown (D2), hot carrier injection (D3), bias temperature instability (D4) and stress migration (D5) have no sample size, lot count or acceptance figure in the table. The data, method and internal criteria need not be repeated for each device, but should be available on request for new technologies. This group is the closest to slow end-of-life failure. Section 4.3 adds that a product qualification is typically not enough to release a new technology for automotive use, and it strongly recommends a separate technology qualification first.

Group E: electrical verification

  • Pre- and post-stress test (E1). Readouts run to the limits of the device specification.
  • ESD (E2, E3). The human body model target is 2 kV, or 1 kV for 28 nm and below or RF pins. The charged device model target is test condition 750 on corner pins and 500 elsewhere, or 250 for those same advanced or RF cases. Parts below target need specific user approval.
  • Latch-up (E4). Three parts from 1 lot, tested at the maximum ambient for the grade under method Q100-004.
  • Electrical distribution (E5). 30 parts from each of 3 lots, or a skewed process lot, across room, hot and cold. It shows that the spread fits the limits.
  • Fault grading (E6). Method Q100-007 asks for at least 98% stuck-at coverage on digital blocks in production, or 97% with acceptable IDDQ testing, and calls 80% a reasonable target for transition delay coverage.
  • Case by case (E7 to E12). Characterisation applies to new technologies and part families. EMC, short circuit and soft error tests run by agreement between supplier and user, and soft error rate applies to parts with 1 Mbit or more of SRAM or DRAM. Lead-free parts follow AEC-Q005.

ESD, latch-up and electrical distribution are the three that must use device-specific data. Generic data from a family is encouraged almost everywhere else, and it has no time limit, so its age is worth asking about.

Groups F and G: screening and cavity packages

Group F is part average testing and statistical yield analysis. The supplier sets the samples and limits, and must run some variant that meets the intent of the guidelines, or justify why it cannot. AEC-Q001 Rev-D describes the aim as removing parts with abnormal characteristics. Group G applies mainly to hermetic and cavity packages: mechanical shock, vibration, constant acceleration, leak, lid torque, die shear and internal water vapour, plus a 1.2 m drop test for MEMS cavity devices only.

An illustrative HTOL calculation

This calculation is illustrative. The mission profile comes from Table A7.1 of the standard, which AEC itself says is a reference example and not an absolute use condition. Your profile should come from your customer.

The example assumes 12,000 hours of operation over 15 years at an average junction of 87 °C. The Arrhenius model is:

AF = exp[(Ea / k) × (1/Tuse - 1/Tstress)]

Take Ea = 0.7 eV, k = 8.61733 × 10-5 eV/K, Tuse = 87 °C = 360.15 K and Tstress = 125 °C = 398.15 K.

  1. Ea / k = 0.7 / 0.0000861733 = 8,123 K.
  2. 1/360.15 - 1/398.15 = 0.00026500 per K.
  3. 8,123 × 0.00026500 = 2.153, so AF = e2.153 = 8.61.
  4. Test time needed = 12,000 / 8.61 = about 1,390 hours, in line with the 1,393 hours shown in AEC's table.

So on these inputs a 1,000-hour test with a 125 °C junction covers only about 8,610 use hours of the 12,000. Now change one input at a time, with hours rounded to the nearest 10:

Illustrative caseEaUseStressAFHours for 12,000 use hours
Base case0.7 eV87 °C125 °C8.611,390
7 °C more junction in test0.7 eV87 °C132 °C12.25980
Lower activation energy0.5 eV87 °C125 °C4.652,580
Higher activation energy1.0 eV87 °C125 °C21.66550
Hotter use, hotter test0.7 eV100 °C132 °C5.582,150

Three lessons follow. First, 7 °C of self-heating in the oven, an assumed figure here, moves the result from short of the target to just past it. This is why the standard's HTOL notes want the test's junction temperature available, not only the ambient. Second, the activation energy changes the answer more than most inputs do. AEC calls 0.7 eV a typical value and gives a range of -0.2 to 1.4 eV depending on the mechanism. Third, a part that runs 13 °C hotter in use needs far more stress time, even with the hotter test.

This model covers temperature alone. It has no voltage term, and it says nothing about cycling or humidity, which have their own models. The standard also warns that going well beyond its conditions can create failures that would never occur in use, so raising the oven temperature is not a free fix.

What a pass leaves open in the field

Appendix 7 is clear about statistics. Zero failures in 231 parts is an LTPD of 1, or at most 1% failures at 90% confidence, and the check is simple: 0.99231 = 0.098, just under 10%. AEC adds that this sample is not sufficient or intended for PPM evaluation.

ELFR is larger, but the same logic holds. As our own illustrative arithmetic, zero failures in 2,400 parts gives an upper bound of 1 - 0.11/2400 = 0.096% at 90% confidence, or about 959 parts per million. Even the biggest sample in the standard cannot see a defect rate in the low hundreds of parts per million. That job falls to process control and to the Group F screens.

Three kinds of field failure sit largely outside the qualification:

  • Test escapes. Coverage figures are tied to fault models, and Q100-007 itself says its theoretical field reject estimates account only for stuck-at faults. Our post on silent data corruption shows how parts that pass factory screening can go on to give wrong results in service. Its evidence comes from data centre processors rather than cars, but the lesson about weak fault models carries over.
  • System conditions. A qualified IC can still sit in a design that fails on supply dips, poor decoupling or a state machine with no way back. Our post on why first hardware designs fail in the field covers that side.
  • Mission profile mismatch. Long service life, high humidity at moderate temperature or a new material can all fall outside the standard's assumptions. Appendix 7 then points to a profile-specific plan or to robustness validation.

Changes also reopen the question. Section 3.2 requires requalification when a change could affect form, fit, function, quality or reliability, and Table 3 guides which tests to repeat. Our automotive and safety-critical lifecycle on The Semiconductor Hub shows where these steps fall in a programme.

Seven questions for a supplier

  1. Which revision of AEC-Q100 is the qualification written against, and what grade?
  2. Which results come from this device, and which are generic family data? How old is that data, and is the family reasoning sound?
  3. What was the junction temperature during HTOL, and what activation energy supports any claim beyond 1,000 hours?
  4. Does the standard duration cover our mission profile, and can we see the calculation?
  5. What are the wear-out results for this technology, and what failure rate do they imply at our temperatures?
  6. What are the ESD levels, including any pin exceptions, and has anything needed user approval?
  7. What part average testing and yield screens run in production, and what stuck-at and delay coverage does the production test reach?

Sources

The HTOL cases and the ELFR bound are illustrative, using the stated inputs. This post summarises the standard in our own structure and does not replace it. It is not legal or compliance advice, so check the full document and your customer's requirements before relying on any condition here.

Murugavel Ganesan
About the Author
Murugavel Ganesan
Product Director | SerDes · IVN · Edge AI · Compute | Ex Intel · Infineon

A semiconductor product leader working across product architecture, definition, strategy, and silicon and software execution, from concept through volume production. He has worked with teams across Asia, Europe and the USA, including at Intel, Infineon and ArrayComm.

His work covers ASICs, SoCs, FPGAs and chiplets for wireless, wireline, automotive, IoT and AI, including PCIe, CXL and DDR interfaces alongside RF and mixed-signal design.

Share Facebook X LinkedIn WhatsApp Email

Post a Comment

0 Comments

Your comments will be moderated before it appears here.

Post a Comment (0)

#buttons=(Ok, Go it!) #days=(20)

Our website uses cookies to enhance your experience. Learn more
Ok, Go it!