What is burn-in, and why are chips "burned in" before they ship?

Burn-in is running a chip powered for a while before it ships, somewhere hotter than normal and at a slightly higher voltage than normal. The point isn't to destroy the chip. It's to make the chips that are weak from birth, the ones that would fail soon after they go into use, fail at the factory and get pulled out, instead of failing in someone's hands.

Weren't the chips already tested? Why would they still fail?

Every chip is tested before it ships, usually twice: once on the uncut wafer, and again in final test after packaging (what is tested and how is in "How are chips tested, and how is wafer sort (CP) different from final test (FT)?"). But a test asks "is it good right now?"

Some chips are good right now but carry a tiny flaw. A fine wire might not be bonded quite firmly enough (how the wires go on is in "What is wire bonding, and how does a chip connect to the outside?"), or a speck of dirt might have got in while it was made. At first everything works, and every test passes. After a few days or weeks of use, the little flaw grows, and the chip fails.

Failing not long after going into use is called early-life failure, sometimes "infant mortality". It usually comes from a small flaw left over from manufacturing, not from a bad design.

It's like a new shirt with one loosely sewn button: you can't tell when you try it on, and it falls off after a few wears.

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 as a line. The shape looks like a bathtub, so it's called the bathtub curve. It has three parts:

1. Early-life failure: when the chips are new, the share that fails is higher, but it drops fast. The ones carrying small flaws fail early, one by one, and the ones left are healthier. This part usually lasts the first few weeks to a few months.
2. Random failure: then comes a long stretch where the share that fails is low and flat. Chips still fail now and then, but it's more like bad luck: there's no telling which one, or when.
3. Wear-out: after many years of use, the materials slowly wear, and the share that fails climbs again.

The tub is high on the left, flat on the bottom and turned up on the right. Burn-in deals only with the left end. It can't help with the bad luck in the middle or the wear-out on the right.

What does burn-in actually do?

Many of the chemical reactions that make a chip go bad run faster the hotter it gets, and a slightly higher voltage pushes weak spots to show themselves sooner. So the chips go into something like an oven, heated and powered, and run for a while. The few that would have lasted some weeks in a user's hands fail at the factory instead.

The temperature is commonly above 100 °C, for example 125 °C. How long depends on the product, anywhere from a few hours to a few days. For the strictest military and space parts, the standards call for burning in every single chip, and for several days.

After burn-in the chips are tested again, and the failed ones are pulled out. The ones that make it through have already passed the left end of the tub, and go on to ship.

Why isn't every kind of chip burned in?

Burn-in sounds great, but it has a cost:

• Time: every batch waits a few more hours to a few days, and shipping slips with it.
• Money: it takes dedicated equipment that runs many chips at once, every one of them powered; it takes up space and uses power.
• The failures are thrown away: the chips that burn out had counted as good, so they come off the yield.
• Too long is bad too: good chips use up a little of their life as well. Overdo it, and you push them closer to the right side of the tub, early.

So not every kind of chip has every unit burned in. It's more common where a failure is very costly: chips in cars, in medical devices, in aircraft and spacecraft, and in equipment like telecom base stations that has to stay on. A failure halfway down the road or in a hospital costs far more than a few more days in the oven.

There's a more basic fix too: find which step the small flaws come from, and fix the process. Once early-life failures are rare enough, burn-in can be shortened, or phased out over time.

If not every chip is burned in, how do you know a batch is good?

One compromise is sampling: take a few hundred chips from a batch, run them powered at high temperature for a day or two, and see how many fail. If many fail, the batch may have a problem, and it needs looking into first. If none fail, you can feel more at ease.

But sampling can't catch every weak chip: the ones not picked still ship as they are. It answers "is this batch roughly good?", not "is this chip good?"

How is burn-in different from HTOL?

Among reliability tests, one often gets mixed up with burn-in: HTOL, high temperature operating life. Both are "heat, voltage, run for a while", but they answer different questions.

Burn-in is screening: every chip that will ship is burned in, for a short time, to pull out the weak ones.

HTOL is qualification: before a new product or a new process goes into mass production, a sample of chips runs nonstop for 1,000 hours, about six weeks, which under common assumptions squeezes years of use into that time. A common standard is three lots of 77 chips each, with not a single failure. What it proves is that this kind of chip, made this way, lasts long enough.

	Burn-in	HTOL
Purpose	Screening: pull out the weak ones	Qualification: is this kind of chip durable enough?
Which chips	Every chip that will ship	A sample
How long	A few hours to a few days	1,000 hours, about six weeks
When	In production, before shipping	Before mass production starts
What counts	Failed ones are pulled, the rest ship	Not a single failure, or it doesn't pass

In one line: burn-in asks "is this chip weak from birth?"; HTOL asks "will this kind of chip last for years?"

In fab24hr

In chapter 12 of the main story, The life of a chip, "The Night Before Shipping", you are the customer liaison at a test-and-assembly house. The first batch of Glowmere, a hearing-aid chip, ships at eight tomorrow morning. Fang in the reliability lab says he can take a few hundred, put them in the oven and run them powered at high temperature for two days to check for early-life failures; on the wall behind him hangs a bathtub curve. But with the analysis, the whole batch would ship a week late. Run the sample first or ship as planned: the decision is yours, and what the sample finds depends on whether the wire bonder was stopped and retuned in the chapter before.

Chapter 4, "Two Weeks Early", is on the R&D line, where you are the R&D integration engineer. HTOL has just passed its 500-hour readout: three lots of 77 chips each, zero failures, three more weeks to go. The customer wants the chips two weeks sooner, and you recommend either waiting for it to finish or going ahead with a risk start.

Tap "Clock in", choose "A day in the fab", tap "Test-and-assembly house" on the "Your role today" card and swipe to the customer liaison: the evening shift is chapter 12. In the morning shift, Fang works with you in the reliability lab on a sample a customer sent back.

In the game, Glowmere's first batch gets an early-life failure sample, not burn-in of every chip. The test-and-assembly house, the chip, the people, and numbers like "two days" and "a week late" are fictional settings; what burn-in, the bathtub curve and HTOL are follows public, general sources. For other jargon, see Chip fab jargon in plain English.