What is reliability testing, and how can you tell in weeks that a chip will last for years?

A reliability test asks: after years of use, will this chip still be good? An ordinary test asks "is it good right now?"; reliability asks "will it still be good in a few years?" The way to find out is to put chips somewhere hotter than normal, at a higher voltage, through hot-cold cycles or into heat and humidity, so that aging that would take years happens within a few weeks, and then work back to how long they would last in normal use.

The chips passed their tests. Why test reliability too?

Every chip is powered up and tested before it ships, usually twice: once on the wafer (wafer sort) and again after packaging (final test); see "How are chips tested?". But chips age. Inside, things change little by little:

• Metal atoms in the wires get pushed along by the current, and over time a wire can thin out or even break.
• The very thin insulating layer in a transistor is under voltage all the time, and over time it can break down.
• The chip, the plastic package and the solder joints expand and shrink by different amounts with heat, so they pull on each other every time things heat up and cool down.

These take years to cause trouble, but nobody can wait years before selling a chip. It's like testing how long a shoe lasts: you don't wear it for three years, you put it in a machine that bends it over and over.

The bathtub curve: when are chips most likely to fail?

Put a large batch of identical chips into use, count how many fail in each stretch of time, and draw it. The shape looks like a bathtub, so it's called the bathtub curve:

1. When they're new: more fail, mostly from small flaws left over from manufacturing. This is early-life failure, and it drops off quickly.
2. A long stretch in the middle: very few fail; when one does, it's more like bad luck.
3. After many years: the materials slowly wear and age, and failures climb again.

Reliability testing looks at both ends: whether the left end is high, and whether the chips last long enough before the right end turns up. There's more on the curve in "What is burn-in?".

How can a few weeks tell you about years from now?

By speeding things up. Many of the chemical reactions that age a chip run faster the hotter it gets, the way milk left on the counter goes bad faster than milk in the fridge. A little more voltage makes weak spots show up sooner, and humidity and hot-cold cycles work the same way. So the conditions are made harsher than normal, and aging happens early.

But you can't push it without limit. Too hot, and failures can appear that would never happen in normal use, such as materials going soft, and then the results say nothing about real use. So the conditions are chosen to speed up the kinds of failure that do happen in normal use. How much faster also depends on the kind of failure: each one reacts to temperature differently.

How is a reliability test done, step by step?

Step 1: Pick the samples

Take a few dozen chips from each of several lots, made and packaged the way mass production will make them. Every lot comes out a little different, and a sample from one lot might happen to be the best one, so the result would only speak for those few chips.

Step 2: Put them under accelerated conditions

The chips are plugged into boards and pushed into an oven or a temperature-and-humidity chamber, then powered, heated and left to run, so that aging happens early. Without speeding things up, you would really have to wait years.

Step 3: Take them out and measure them every so often

Every few hundred hours, say, the chips come out and are measured one by one: does each still work, has its leakage gone up? Then they go back in. A full test often takes over a thousand hours, several weeks. Measuring along the way shows when a chip started to go bad and whether it drifted slowly; looking only at the last measurement can't tell you that.

Step 4: See how many failed, and how

A common pass mark is zero failures. If any fail, they are opened up and cut apart to find out why (see "What is failure analysis (FA)?"). The cause may be in the process, the design or the package, and each needs a completely different fix; without finding it, you might fix the wrong thing.

Step 5: Work out how long they'd last in normal use

Using the rule "the hotter, the faster the reaction", the time in the oven is converted into time in normal use, to answer "will it last for years?" Zero failures isn't a guarantee, either. It lets you say how sure you are that the failure rate is below a certain level; with too few samples or too short a test, you can't say it with much confidence.

What are the common kinds of test?

• High-temperature operation (HTOL, high temperature operating life): over 100 °C, with voltage applied, the chips keep working for more than a thousand hours. It tests whether the transistors and wires slowly age, because a chip is powered and warm every day.
• Temperature cycling: back and forth between tens of degrees below zero and over 100 °C, many times over. It tests whether the joints inside the package crack, because materials expand and shrink by different amounts; every time a phone goes from an air-conditioned room into the hot sun, they pull on each other once.
• High temperature and humidity: in the 80s °C, at over 80% humidity, with voltage applied, for more than a thousand hours. It tests whether moisture gets through the plastic package and corrodes the metal wires, because plastic isn't completely waterproof.
• Static electricity: a discharge is fired into the pins, imitating a person's hand touching them, or the chip itself picking up a charge and then touching metal. It tests whether the protection circuits in the chip can stop it, because a person can easily carry static of over a thousand volts, and chips get touched all the way from the factory to assembly.

Which tests to run and how many chips to take are set out in public industry standards, and also depend on where the product will be used: chips that go into cars are tested more, and more strictly (see "The news says "automotive chips"").

Why take extra care with the lots that had very low yield?

Very low yield often means those wafers or lots were hit by especially many defects. Some defects aren't big enough to kill a chip on the spot; they leave a small flaw that passes the tests but makes the chip weaker. Studies have found that in wafers and lots with very low yield, a higher share of the chips that passed their tests fail early later on.

It's a statistical tendency, not a sign that every chip has a problem. But demanding products give these an extra screen, such as burn-in, or handle them separately (see "Why do bigger chips have lower yield?").

How is reliability testing different from burn-in, and when is it done?

Reliability testing is a sample: a few chips are tested to find out how long this kind of chip, made this way, will last. Burn-in is every chip: before shipping, each one is burned in for a while to pull out the ones that are weak from birth and would fail early. Think of shoes: reliability testing wears a few pairs down until they fail, to see how long this model lasts; burn-in has every pair walked in for a bit first, so the ones whose soles aren't stuck on properly get pulled out.

Reliability is mostly tested before mass production. A new chip first has a pilot run on the R&D line; at the transfer review before it goes to the production fab, reliability is one of the items to check (see "What does an R&D integration engineer do?" and "How long does a new chip technology take to go from a paper to mass production?"). Starting wafers in the production fab before the tests are done is called a risk start.

In fab24hr

Tap "Clock in", choose "A day in the fab", tap "R&D line" on the "Your role today" card and swipe to the R&D integration engineer. The evening shift is the night before the transfer review. On the eight-item checklist, "Reliability (HTOL)" has just passed its 500-hour readout: three lots of 77 chips each, zero failures, three more weeks to go. Hsiao-Wen in the electrical test lab says: "In my experience, if nothing fails in the first 500 hours, nothing usually fails after. But usually isn't always." The customer for Glowmere, a hearing-aid chip, wants the chips two weeks sooner; you recommend either waiting for HTOL to finish or going ahead with a risk start. This night is chapter 4 of the main story, The life of a chip, "Two Weeks Early".

The reliability lab at the test-and-assembly house has ovens and a bathtub curve on the wall; in chapter 12, "The Night Before Shipping", Fang, the reliability engineer, offers to take a few hundred chips and run them for two days to check for early-life failures. Under "Learn" on the home screen, tap "All 31 machines" to take the reliability oven apart and see how the sampled modules are plugged into boards and pushed in to be heated and powered.

The fabs, Glowmere, the people and numbers like "500 hours" and "three weeks" in the game are fictional settings; what reliability testing is and how it's done follow public, general sources. For other jargon, see Chip fab jargon in plain English.