Semiconductor yield is the share of chips on a wafer that work. A wafer is a thin, polished disc of silicon, usually 300 millimetres across, on which a factory makes hundreds of chips at the same time. A fab, the factory that makes chips, earns money only on the chips that work. So yield decides whether a fab makes a profit, and finding defects early, and fixing their cause, is worth a great deal. Yield is both a technical measure and a business one.
Here is the technical side. The fab builds the chips on the wafer layer by layer, through hundreds of steps over about three months. Each chip on the wafer is a small rectangle called a die. A die holds millions or billions of transistors, tiny electrical switches that are the basic parts of every chip. The smallest features on a leading chip are only a few nanometres wide. A nanometre is a millionth of a millimetre. At that size, one speck of dust, a scratch or a layer that is slightly too thick can break a die. At the end, every die is tested, the wafer is cut up, and each working die is put in a package. A package is the small plastic or ceramic case, with metal contacts, that is soldered onto a circuit board. Dies that fail are thrown away.
Now the business side. The fab spends about the same money on every wafer, whether its chips work or not. Say one wafer costs ₹10 lakh to make, and it holds 600 dies. If 90 per cent of them work, there are 540 good chips, and each one costs about ₹1,850. If only 60 per cent work, there are 360, and each one costs about ₹2,800. Large dies suffer most, because each one covers more area and so has a greater chance of catching a defect. A new manufacturing process usually starts with a low yield. The fab's engineers raise it over months by finding and removing the causes of defects.
Take, for example, wafer inspection and measurement. Its machines check each layer for defects too small to see, so the fab can correct a faulty step in the manufacturing equipment before it spoils thousands of wafers.
Consider a wafer-scale chip, which uses a whole wafer as one chip. A chip that large is certain to contain some defects, so its designers add spare cores and send data around the broken ones.
Imagine a plant for chip assembly and test. It tests every packaged chip again, and a chip that fails at this stage is a larger loss, because its package has already been paid for.
How is yield different from engineering tolerances? A tolerance says how far a part may differ from its drawing and still be accepted. Yield counts how many of all the parts made end up accepted. The tighter the tolerances, as on a modern chip, the harder a high yield is to reach. A foundry and its customer agree in their contract who pays for the chips that fail.
Appears in
- Wafer-scale chip↗ AI and data centre chipsA wafer-scale chip uses a whole wafer as one chip, so it is certain to contain defects, and its designers add spare cores to work around the broken ones.
- Semiconductor manufacturing equipment↗ SemiconductorsDeposition and etch machines lay down and remove each layer of a wafer, and a layer that is too thick or too thin breaks the dies built on it.
- Wafer inspection and measurement↗ SemiconductorsWafer inspection systems find defects on each layer while the wafer is still being made, so the fab can correct the faulty step before more chips are lost.
- Semiconductor foundry↗ SemiconductorsA foundry earns only on the working chips from each wafer, so raising the yield of a new manufacturing process is one of its main jobs.
- OSAT (chip assembly and test)↗ SemiconductorsAn OSAT plant tests every die before packaging and every packaged chip after it, and throws away the chips that fail.