What Is Moore's Law? Is It Dead? A Plain Explanation
Moore's law is a rule of thumb: the number of transistors on a chip doubles about every two years. Gordon Moore wrote "every year" in 1965 and changed it to "every two years" in 1975. It is not a law of physics. It is a forecast drawn from past trends, and it also became a target for the industry. Transistor counts are still rising. But the cost per transistor falls more slowly than before, and speed gains have slowed too.
What you'll learn
- What Moore actually wrote in 1965 and 1975
- Why transistor counts kept rising, and why each transistor got cheaper
- Why shrinking also made transistors faster and more power-efficient (Dennard scaling), and why that stopped
- What the "Is Moore's law dead?" debate is really about, and where things are heading
Prerequisites: it helps to know that a transistor is "a very small switch that turns electric current on and off." "What Is a Transistor?" explains this.
In everyday life: Moore's law in the news
"The limits of Moore's law." "Moore's law is over." You often see headlines like these in chip news. In 2022, Jensen Huang, the CEO of NVIDIA, said "Moore's Law is dead." NVIDIA is known for its AI chips. Yet in 2023, Intel's CEO at the time said the pace had slowed, but it was not over.
The two statements sound like opposites. In fact, they are counting different things. This article starts with what Moore actually wrote and goes step by step from there.
Today's smartphone chips hold billions to tens of billions of transistors. Intel's 4004 from 1971 had about 2,300. Moore's law describes this growth.
What Moore wrote in 1965
Gordon Moore was an engineer who later co-founded Intel. In 1965 he was working at a company called Fairchild. He wrote a short article for the April 19 issue of the magazine Electronics. It was titled "Cramming more components onto integrated circuits." An integrated circuit (IC) is a circuit whose parts, such as transistors, and wiring are built together on one small piece of silicon.
Moore cared less about the number of components itself and more about the cost per component. A chip with too few components is costly per component, because costs such as packaging it into a product are spread over only a few. A chip with too many has more faulty units, which also raises the cost. So there is a "just right" number of components where the cost per component is lowest.
Moore wrote that this lowest-cost number of components had roughly doubled every year. In his words: "The complexity for minimum component costs has increased at a rate of roughly a factor of two per year." At the time, that number was about 50 components per chip. If the trend continued, he predicted, it would reach about 65,000 by 1975.
It was a bold forecast. He looked ten years ahead from only a few years of data. Note that he counted "components," which included resistors and other parts, not only transistors.
The 1975 revision: "doubling every two years"
Ten years later, in 1975, Moore had moved to Intel. He gave a talk at IEDM, an international conference on electronic devices. According to the Computer History Museum, his 1965 forecast had turned out to be about right.
In this talk, Moore gave four reasons for the growth so far. The first was the technique of copying patterns with light. The second was the size of the wafer, the round silicon plate that chips are made on. The third was the manufacturing process. The fourth was what he called "circuit and device cleverness," meaning smarter design of circuits and devices. He then slowed the pace of his forecast to, in the museum's words, "a doubling every two years, rather than every year." The "doubling about every two years" you often hear today is this revised version.
Moore did not name "Moore's law" himself. The Computer History Museum says the name came from Carver Mead, a professor at Caltech (the California Institute of Technology). The glossary also has a short definition of Moore's law.
Checking with a chart: did it really double every two years?
Let's plot the numbers for real chips. They grow tens of millions of times over, so they are hard to fit on one chart. Here, each step up the vertical axis means 1,000 times more. This is called a logarithmic (log) scale. On this scale, "doubling at a steady interval" shows up as a straight line.
Over the 52 years from the 4004 to the M3 Max, the count grew about 40 million times. That is about 25 doublings, so the count doubled about every two years on average (our calculation). The dots sit close to a straight line. This shows that the 1975 outlook held true for a long time.
Not a law, but an observation and a target
It is called a "law," but it is not a rule of nature like the law of gravity. It is an observation of past trends, and a forecast based on them.
One reason it held for so long is that the industry made this pace its goal. From 1992, an industry group drew up an outlook of how fine chips would be made, and by when. This kind of plan is called a roadmap. From 1998 it became an international roadmap (the ITRS). Equipment and materials companies planned their development to match it. The forecast became a target, and the target helped the forecast come true.
How it works 1: three ways the count went up
The number of transistors on a chip comes from two things multiplied together: how small each one can be made, and how large the chip can be. On top of that, clever designs do the same job with fewer parts.
Making things smaller
The biggest factor was shrinking transistors and wiring. This trend is called "scaling." The fine patterns on a chip are made by using light to copy a master pattern. Light with a shorter wavelength can copy finer patterns. "What Is EUV Lithography?" explains the most advanced light used today.
In the past, the rule of thumb was to shrink lengths to about 0.7 times with each generation. If both width and height become 0.7 times, the area is about half. So about twice as many transistors fit in the same area. To learn how to read generation names like "3 nm" and "2 nm," see "What Does 2nm or 3nm Mean?"
Making chips larger
The area of each chip also grew little by little. In Moore's 1975 paper, the area of the most complex chips grew about 20 times from 1959 to 1975. But bigger chips mean more faulty ones. There is also a limit to the area that can be printed at one time. "What Is a Chiplet?" explains a way around this wall.
Clever circuit design
Designs that do the same job with fewer parts also helped. In memory, for example, engineers cut the number of parts needed to store one bit (a single 0 or 1). In 1966, Robert Dennard of IBM came up with a way to store one bit with one transistor and one part that holds electric charge. The patent dates from 1968. The working memory (DRAM) in today's PCs and smartphones uses this method.
How it works 2: why each transistor got cheaper
The cost of processing one wafer goes up with each new generation. But if the number of transistors made from one wafer grows even faster, the cost per transistor goes down. For a long time, this held true.
Wafers themselves also got bigger. The larger the diameter, the more chips you get from one round of processing. Mass production on 300 mm wafers began around 2001 (see "What Is a Silicon Wafer?").
A 2018 study by the economist Kenneth Flamm looked at this. While scaling went smoothly, the cost of making one transistor fell by 20–30% a year. That is why the performance you could buy for the same price kept rising.
How it works 3: smaller also meant faster and lower power (Dennard scaling)
If only the count went up, you might expect chips to get hotter and hotter. But for a long time, shrinking also made each transistor faster and less power-hungry. A 1974 paper by Robert Dennard of IBM and his colleagues showed why.
Here is the main idea in words. Shrink all the dimensions of a transistor by the same factor, and lower its voltage by the same factor. Then each transistor switches faster and uses less power. On top of that, the power per unit area, in other words the heat, stays the same. This relationship is called Dennard scaling.
Thanks to this, each new generation gave a chip of the same size more transistors that were faster, without getting hotter. A chip's circuits work in step with a signal that arrives at a steady rate. The number of times this signal arrives per second is called the clock frequency. The 4004 ran at about 0.75 MHz (about 750,000 times a second). By 2004, Intel's PC CPUs had reached 3.8 GHz (3.8 billion times a second).
| While Dennard scaling held | From the mid-2000s | |
|---|---|---|
| Dimensions | Shrink each generation | Keep shrinking |
| Voltage | Lowered along with size | Can hardly be lowered |
| Heat per unit area | Stays the same | Tends to rise |
| Clock frequency | Rises each generation | Stalls at a few GHz |
To keep things simple, the table shows only the broad trends. The actual turning point differs by a few years from product to product.
The end of Dennard scaling: what happened in the mid-2000s?
Around the mid-2000s, it became hard to lower the voltage any further. A transistor has a voltage at which its switch starts to turn on. This is called the threshold voltage. To lower the supply voltage, this voltage must come down too. But a lower threshold voltage makes transistors leak much more current, even when they should be off.
If you keep shrinking without lowering the voltage, the heat per unit area keeps rising. Raising the clock frequency adds even more heat. There is a limit to how much heat can be removed from a chip. So the frequency could no longer keep going up.
In 2004, Intel cancelled a planned 4 GHz Pentium 4 (according to reports at the time). Instead, it shifted its effort to products with two computing centers on one chip. Each of these centers is called a core. A 2011 report by the US National Research Council also concluded that single-processor performance growth slowed in the early 2000s. "What Is a CPU?" explains cores.
The key point is this: what ended was Dennard scaling, not Moore's law. Transistor counts kept rising after that. What disappeared was the free bonus that more transistors would also mean faster, lower-power chips.
Is Moore's law dead?
The answer depends on who you ask. There are three main views.
The count is still rising: According to Apple, its M1 PC chip (2020) has 16 billion transistors, and the M3 Max (2023) has 92 billion. NVIDIA announced its B200 AI chip in 2024. It joins two chips into one product, with 208 billion transistors in total.
The pace has slowed: In 2023, Intel's then-CEO Pat Gelsinger said the doubling time was now closer to three years. Even so, his position was that Moore's law is not over.
In terms of cost, it is over: At the leading edge, both the tools and the process steps cost more, so the cost per transistor is hard to bring down. According to Tom's Hardware, a Google engineer showed at IEDM 2023 that this cost has stayed roughly flat since the 28 nm generation. But this is one company's view, and costs also vary with the product and the production volume. The NVIDIA CEO's remark at the start was also an answer to a question about GPU prices.
So the answer depends on what you count. Transistors per chip are still rising, though at a slower pace. If you mean the cost per transistor, or the speed of each transistor, progress is not what it used to be.
That is why the industry now combines shrinking with other paths to raise performance.
- Changing the transistor's shape: the part that acts as the switch (the gate) wraps around the current's path on all sides (GAA)
- Changing how wiring is routed: moving the power wiring to the back side of the wafer (backside power delivery)
- Splitting and joining: making a large chip as several smaller chips, then combining them into one product (chiplets)
- Stacking: stacking chips vertically and connecting them (3D packaging). For an example, see "What Is HBM?"
The essay "Beyond Scaling" gives an overview of these directions.
Another direction: "More than Moore"
So far, the story has been about making computing circuits smaller and more numerous. The industry roadmaps (the ITRS and its successor, the IEEE's IRDS) call this "More Moore." Another direction adds value with functions such as sensing or power control. This is called "More than Moore." For these parts, materials and methods that suit the job matter more than the finest patterns. Examples include power semiconductors and image sensors.
Common misconceptions
Myth: Moore's law says computer speed doubles every 18 months
In fact, Moore wrote about the number of components (later, transistors). His pace was "every year," and later "every two years." Moore himself has said he never mentioned 18 months. The 18-month figure restates the trend in terms of performance. It combines the rising count with each transistor getting faster. This restatement is said to have come from David House, then an Intel executive. Speed gains slowed sharply from the mid-2000s, when Dennard scaling stopped.
Myth: It is a law of physics, so it will go on by itself
In fact, it is a forecast based on past trends, and it was also an industry target. It has continued because companies spent huge sums to rebuild their equipment, materials and designs. The word "law" easily makes it sound like a rule of nature.
Myth: Transistors shrink at the same pace as the "X nm" numbers
In fact, "3 nm" and "2 nm" are names for generations. They are not the size of any part of the chip. Today, density rises not only by shrinking dimensions. New shapes help too, such as standing transistors on end or stacking them, and so do smarter layouts. How much the number shrank does not match how much the count per area grew. "What Does 2nm or 3nm Mean?" explains more.
By the numbers
| Number | What it is |
|---|---|
| About 2,300 | Transistors in Intel's 4004 (1971) (Intel) |
| 92 billion | Transistors in Apple's M3 Max (2023) (Apple) |
| About 40 million times | Growth over the 52 years from the 4004 to the M3 Max. Average doubling time about 2 years (our calculation) |
| About 0.75 MHz → 3.8 GHz | Clock frequency of the 4004 (1971) and of Intel's PC CPUs in 2004. About 5,000 times higher |
| 20–30% a year | How fast the cost of making one transistor fell while scaling went smoothly (Flamm, 2018) |
When you see Moore's law in the news, check which number is being talked about.
Learn more
- An overview of paths other than shrinking: the essay "Beyond Scaling"
- How transistor shapes changed, in tables: "Planar MOSFET vs. FinFET" and "FinFET vs. GAA"
- Power used for switching vs. power lost to leakage: "Static vs. dynamic power". To try changing the voltage and frequency, use the CMOS power calculator
- Why some work does not speed up even with more cores: the Amdahl's law calculator
- Related articles: "What Is an Integrated Circuit?" and "What Is a Semiconductor Foundry?"