What is Moore’s law?

On April 19, 1965, the magazine Electronics published Gordon Moore’s paper “Cramming More Components onto Integrated Circuits.” Moore was then one of the founders of Fairchild Semiconductor, a chip company, and had been director of its research laboratories since 1959 (he cofounded Intel later).

An integrated circuit is a set of parts, such as transistors, built together on one small chip. A transistor is a part that lets electricity flow or stops it, like a tiny switch. The paper explains that for simple circuits, the more parts you put on one chip, the lower the cost of each part. But add too many, and more chips fail to come out right, so the cost goes back up. So at any given time there is a number of parts at which each part is cheapest; in 1965 it was about 50 per chip.

This “cheapest number” had been roughly doubling every year. Since that should continue for at least ten years, the paper forecast that by 1975 it would reach 65,000 parts per chip. In 2008, Moore looked back on it as “a wild extrapolation of very little data,” saying he was just trying to get across the idea that integrated circuits were going to be the route to cheap electronics.

The paper also said that integrated circuits would lead to such wonders as home computers (or at least terminals connected to a central computer), automatic controls for automobiles and personal portable communications equipment.

What Moore said was two years. In 1975, at the IEEE (an engineering society) International Electron Devices Meeting, he revised his forecast from doubling every year to doubling every two years. In 2008 he recalled that over the ten years from 1965, the ten doublings he had predicted turned out to be nine, pretty close.

“Doubling every 18 months” is also often quoted, in the press and elsewhere. But Moore himself said in a 1997 interview, “I never said 18 months.”

The difference in doubling time looks small but adds up. Starting from 1, after ten years: doubling every year gives 1,024 times (2 multiplied ten times), every year and a half about 100 times, and every two years 32 times.

The name “Moore’s law” was given by Carver Mead, who was a professor at Caltech. In a 2015 interview, Moore said, “I had a colleague who saw that and dubbed this Moore’s Law.”

A law of physics is a rule of nature that stays the same whatever people do. Moore’s law isn’t like that. Intel, too, says Moore’s law is not a scientific law (it’s not a natural phenomenon). The number of parts kept growing because engineers kept coming up with new ideas. Intel calls it a projection that relies on innovation and technological advancement for its continued truth.

The forecast also became the industry’s target. Because it was predicted, people worked toward it, and so it came true: it became a kind of “self-fulfilling prophecy” and one of the driving principles of the industry. Intel also describes it as a target for Intel and its competitors to deliver against time and again (this is the company’s description).

In a 2015 interview, Moore himself said the term became something that almost drove the semiconductor industry rather than just recording its progress.

For a long time, speed rose along with the number of parts. As Robert Dennard and his colleagues showed in 1974, if you make the parts smaller and lower the voltage in the same proportion, the parts get faster and use less energy each time these switches flip. So in each generation, chips could double the number of parts and raise the clock rate (how many beats a chip ticks each second to keep its steps in time; more beats means more steps each second) while using the same power. This is called “Dennard scaling.”

But as the parts got smaller, more current leaked through even when it shouldn’t, and the voltage became hard to lower any further. By around 2004, raising the clock rate pushed the power a chip used, and the heat it gave off, toward the limits of what air cooling can handle. So the fast rise in clock rate stopped, and the speed of a single processor began to rise only slowly.

Even so, the number of parts kept growing. So chip companies began putting several processors (cores) on one chip. But more cores only make things faster when the work can be split up and done at the same time, and as of 2011, relatively little software was written that way.

No one knows. There is a limit to how small parts can be made. In 2008, Moore said that any forecast of a physical quantity doubling again and again runs into trouble some day, and that the devices being made were getting pretty close to molecular dimensions, which would become a fundamental limit on shrinking them.

In a 2015 interview, he said, “We won’t have the rate of progress that we’ve had over the last few decades,” and “I guess I see Moore’s Law dying here in the next decade or so” (his own forecast).

Intel, on the other hand, wrote in 2023 that with advanced packaging technology (the way chips are enclosed and connected) and new materials, the pace of Moore’s law continues. The page also quotes an Intel executive: “we will maintain Moore’s Law for the next decade or longer” (the company’s forecast).

Training an AI takes an enormous amount of calculation, and that calculation is done by chips packed with parts. For example, the H100, announced by NVIDIA in 2022, is a computing chip (a GPU) used for things like AI training, and it has 80 billion transistors.

A 2022 study (Sevilla et al.) looked at how the amount of compute used to train AI models has changed over time. It found that before 2010, training compute grew in line with Moore’s law, doubling roughly every 20 months, but that in the deep learning era since around 2010, it has grown much faster, doubling about every 6 months.

As for the singularity, the idea that machines smarter than humans will arrive and the future can no longer be foreseen, Moore said in 2008, “I am a skeptic. I don’t believe this kind of thing is likely to happen, at least for a long time” (more on the “Technological singularity” page).