What does "yield" mean in chip news, and why does 1% matter so much?

When the news mentions chip "yield", it means the share of the chips made on a wafer that come out good. Make 100 chips, get 90 good ones, and the yield is 90%.

Ninety percent sounds fine, yet the news treats a few points of yield as a big story. To see why, start with how chips are "baked".

How many chips are on one wafer?

Chips aren't made one at a time. A factory takes a thin, round slice of silicon called a wafer and builds many identical chips on it at once: a wafer can hold over a thousand small chips, or only a few dozen very large ones. At the end, the wafer is cut apart and each chip is packaged.

It's a lot like baking cookies. A whole tray goes into the oven together, and a few always come out burnt or broken. Yield asks one question: how much of the tray can you sell?

Why can't every chip be good?

The finest lines on a chip are more than a thousand times thinner than a hair. At that size, chips fail in two common ways.

One is something landing on them. A particle too small to see lands in the wrong spot and can join two lines together or cut one in two. However clean the cleanroom is, it never gets to zero.

The other is coming out slightly off. A film is a little too thick, or the line width a little too wide. The chip may still work, but if it runs too slowly or leaks too much current while idle, it doesn't count as good either.

At 99.9% per step, how much is left at the end?

From wafer start to finish, a wafer goes through hundreds of steps in the fab, and any of them can ruin a few chips. Here's a simplified way to look at it: say each step ruins only 1 chip in 1,000, so each step is 99.9% good.

After one step, 99.9% is left. After two, 99.9% × 99.9%. Keep multiplying:

• After 100 steps: about 90% left
• After 300 steps: about three quarters left
• After 500 steps: about 60% left

Each step loses only a little, but multiplied together it adds up. Put the other way: to finish 500 steps with 95% left, each step has to reach 99.99%, only 1 bad chip in 10,000.

In a real fab, the risk is different at every step; this only shows the power of multiplying. Any step that drifts a little gets multiplied into the final yield. That's why every step in a fab is handled so carefully.

Why do bigger chips have lower yield?

Back to the cookies. Say every tray gets 10 burnt spots, scattered at random.

With small cookies, 100 to a tray, 10 spots can ruin at most 10 cookies, so 90% can still be sold. With big cookies, only 10 to a tray, the same 10 spots hit six or seven of them on average.

Chips work the same way. With about the same number of defects, the bigger the chip, the more likely it is to be hit. So before reading a yield number, you need to know how big the chip is: the yields of two chips of different sizes can't be compared directly.

Why does 1% matter so much?

A wafer costs about the same no matter how many good chips come out of it. The materials, tool time, power, water and people are all spent while it's being made. The fewer good chips, the bigger the share of the cost each one has to carry.

Try some made-up numbers. Say one wafer holds 1,000 chips:

Yield	Good chips	Cost each good chip carries
90%	900	1×
60%	600	1.5×
45%	450	2×

At high volume, 1% shows even more. Now say a fab makes 10,000 of these wafers a month (also a made-up number). If yield goes from 90% to 91%, each wafer gives 10 more good chips: 100,000 more a month. Those wafers were going to be made anyway, so the fab spends almost nothing extra for them. The other way round, losing 1% means 100,000 chips made that can't be sold.

A lower yield than planned brings one more problem: there aren't enough chips for what was promised to the customer. Making up the gap means starting more wafers, and a wafer takes weeks to months from start to finish, so the extra chips arrive late.

Why is yield low when a chip is new?

When a new process generation (like the "2 nanometer" in the news) or a newly designed chip is first made, many problems haven't been found yet: a step whose settings aren't quite right, a tool that treats the wafer edge differently, a speck on a photomask. The first few lots of a new product are its pilot run, and the yield then is often low.

Engineers track down the causes one at a time. They look at which kind of failure is most common and where on the wafer it sits, fix the biggest one, then go after the next. Each fix lifts the yield another step.

Toward the end, the climb slows down, as the remaining causes get smaller and harder to find. Once the yield is high enough for the chip to pay off, mass production starts. This climb, which gradually levels off, is what the news calls a "yield ramp"; some call it a learning curve.

How do they know which chips are good?

Before a finished wafer is cut apart, it usually goes through wafer sort: a probe card presses many very fine needles onto the chips' contact pads, powers them up and measures them. Every chip is tested, not just a sample.

Each chip is then sorted by its result: the good ones in one group, the bad ones in a separate group for each way of failing. These groups are called bins. The results are also drawn by position as a wafer map, with good and bad in different colors. After the wafer is cut, only the good chips go on to be packaged.

The wafer map is also a clue to the cause. Failures scattered with no pattern look more like randomly falling particles. Failures bunched in a ring at the edge, in the same spot of every square the scanner prints, or along an arc usually point to a particular step, tool or photomask.

Next time you see it in the news

• "Poor yields": more of the chips came out bad than expected. Each good chip costs more, and the promised quantity may not be ready in time.
• "Yields are improving": the same wafers now give more good chips. Costs come down, and more chips can be delivered.
• "Yield ramp" or "ramping yields": a new process or chip has just started, engineers are fixing the causes of failure one by one, and the yield is climbing step by step.
• "Yields have reached mass-production levels": the share of good chips is high enough to pay off, so large-scale production can start.

In fab24hr

In chapter 3 of the main story, The life of a chip, you're the R&D integration engineer on the R&D line. The first lot of "Glowmere", a hearing-aid chip, has been through wafer sort, and each of three wafer maps shows a different odd shape (a ring at the edge, a repeating grid, an arc). You work out where each one points. In chapter 10, you're the test engineer at the test-and-assembly house: you find the most common failure in the bins, see where it sits on the wafer, and decide where to draw the limit for standby current. Tighten it, and the yield falls by almost five percentage points and the first batch comes up short. The companies, chips, people and yield numbers in the game are fictional; what yield is and why it never reaches 100% come from public, general sources. For other jargon, see Chip fab jargon in plain English.