What is Moore's law, and why do people say it's slowing down?

Moore's law says that the number of transistors on a chip roughly doubles every so often. It isn't a law of physics. It's an observation that Gordon Moore, an engineer working on chips, drew from the data in 1965: at first he said about every year, and in 1975 he changed it to about every two years. When the news says it's slowing down or even dead, it's because transistors are now close to the size of atoms and keep getting harder and more expensive to make, so the doubling no longer keeps up with every two years.

What did Moore actually say?

In 1965, Moore wrote a short article for an electronics trade magazine. Integrated circuits had been invented only a few years earlier, and the most complex chip held only about 64 components (transistors, resistors and so on).

He plotted the numbers from the previous few years. There were only five points, but they showed the count doubling about every year. So he boldly extended the line ten years: by 1975, a chip would hold about 65,000 components. Doubling every year for ten years is ten doublings, about a thousand times; 64 times a little over a thousand is about 65,000.

What he was really looking at was the number that made the most economic sense: fit a few more on one chip and each component gets a little cheaper on average, but fit too many and more chips come out bad, so the cost goes back up. So from the very start, Moore's law was about money as well as technology.

Why "every two years", and where does "18 months" come from?

In 1975, Moore made a new estimate. He expected things to slow a little and changed the pace to about every two years. That's the version most people mean by Moore's law, and for decades afterwards the industry roughly kept to that rhythm.

The "18 months" you often hear isn't Moore's. Someone else added "each transistor also gets faster" to the count and estimated that chip performance would double about every 18 months, and over time the two got mixed up. Even the name "law" was given later, by other people.

If it isn't a law of physics, what is it?

A law of physics, like things falling when you let go of them, holds no matter what anyone does. Moore's law is different: it's an observation and a prediction, and it only comes true if a lot of people keep working at it.

Later it became a target for the whole industry. The industry wrote technology roadmaps together, setting out roughly which year each generation should reach which size, and everyone planned their research around that schedule. That's why people often call it a self-fulfilling prophecy: everyone believed the count would double, so everyone worked hard to make it double.

Why do smaller transistors make chips faster, more efficient and cheaper?

A transistor is a tiny switch on a chip, run by electricity (What is a transistor? explains it in detail). Making it smaller brings three things:

• More of them: the same space holds more switches, so the chip can do more.
• Faster, on less power: a smaller switch needs less electricity for each flip, and flips faster.
• Cheaper: even if making a wafer costs a bit more each generation, each wafer holds roughly twice as many transistors, so each one costs much less on average.

These three happened together for decades. Doubling every two years means about a thousand times in twenty years and about a million times in forty. In the early 1970s, the first processor chips held a few thousand transistors; a phone chip now holds more than ten billion, millions of times more. That's how phones became so powerful and still affordable.

The "nanometer" names in the news are the names of these generations as they shrank one after another, and today they no longer match any real size (see what "2 nm" really means).

Why do people say Moore's law is slowing down?

From around the 2010s, many people noticed that the doubling had become a little slower than every two years. There are three reasons.

1. Sizes are close to atoms

Silicon atoms sit side by side, and one nanometer holds only a few of them. In the most advanced transistors, some parts are only a few dozen atoms wide, and the insulating layer under the gate once got as thin as one or two nanometers, until a new material was needed. You can't cut an atom in half. At this size, a few atoms more or less can change how a transistor behaves, and there's less and less room left to shrink.

2. Leakage and heat won't come down

A transistor that's too small doesn't shut tightly: even when it's off, a little current slips through. That's leakage. When the insulating layer is too thin, electrons also go straight through it. For a long time, each smaller generation also ran at a lower voltage, so the same area of chip used about the same power. By the mid-2000s, leakage had grown so large that the voltage couldn't come down much more, and the more densely the switches were packed, the hotter the chip got.

That's why the clock speed of computer processors, how many times per second they switch, has barely climbed since then. Instead, one chip got several cores that compute together.

3. Every generation is harder and more expensive

To print finer patterns, the scanner has to switch to light with a shorter wavelength, and the finest patterns sometimes need one layer to be exposed several times. A chip goes through more and more steps, and every step can go wrong. Building a leading-edge fab now easily costs over ten billion dollars. Part of the money saved by shrinking gets eaten up by ever more expensive tools and ever more steps.

So is Moore's law dead?

Opinions differ, and the difference is in the definition. If you strictly mean "doubling every two years, with each transistor getting cheaper", that's hard to do now, so some people say it's dead. If you only mean "each generation fits more than the last", it's still moving, just more slowly and at a higher price, so others say it isn't over. The industry still has no consensus.

How is the industry keeping it going?

Shrinking everything in proportion is no longer enough, so the industry takes other roads. These are the ones usually mentioned:

1. Change the transistor's shape: from flat, to a thin standing wall, to a channel surrounded on all four sides, so even tiny switches shut tightly (covered in What is a transistor?). Researchers are also trying to stack the transistors themselves on top of each other.
2. Build upward: when there's no more ground, build taller. Some memory chips have long stacked their storage cells layer upon layer, and chips can be stacked directly on top of one another too.
3. Put the pieces together with advanced packaging: instead of making everything on one big chip, split it into a few pieces, make each with the generation that suits it best, and put them in one package, close together and densely connected (what "advanced packaging" means).
4. Finer light: the finest layers are printed with light of a much shorter wavelength, extreme ultraviolet (EUV).

And not every chip needs the newest generation: many chips in cars and home appliances do fine with methods several generations old (what "mature nodes" means).

In fab24hr

The fab in the game is fictional and never says how many nanometers it is. The hearing-aid chip Glowmere in the main story, The life of a chip, uses a mature process generation, not the smallest transistors. But the wall that shrinking transistors run into shows up on the R&D line: in A day in the fab, as the furnace R&D process engineer, the morning shift asks how thin to grow the oxide layer under the gate. Thinner lets the transistor switch faster; too thin, and electrons pass straight through and the chip leaks in standby, while Glowmere runs on a single button cell. Every new process goes like this: it's worked out on the R&D line first and handed to the production fab once it's stable. Tap "Clock in", choose "A day in the fab", then on the "Your role today" card tap "R&D line", swipe to the R&D process engineer and tap the Furnace area. The companies, chips, people and numbers in the game are fictional; the Moore's law parts come from public, general sources, and where there's no agreed number, only the order of magnitude is given. For other jargon, see Chip fab jargon in plain English.