How well has Moore's Law held up?

Moore’s Law is the best-known entry in my laws note. It is also the one people most often call dead. So I gathered twenty CPUs, checked each transistor count against a primary or reputable source, and put Moore’s forecast next to them.

What Moore said

In April 1965 Gordon Moore wrote “Cramming more components onto integrated circuits” for Electronics. He predicted that the number of components on a chip would double every year for the next ten years. At the 1975 IEEE International Electron Devices Meeting he slowed that to every two years (Computer History Museum). The two-year version is the one we still quote.

The chart

Each dot is a CPU, from the Intel 4004 (1971, 2,300 transistors) to AMD’s EPYC Turin (2024, about 150 billion). The y-axis is logarithmic, so steady doubling draws a straight line. The orange line starts at the 4004 and doubles every N years; drag the slider to change N. Pick an era to see what let chipmakers keep adding transistors in that period.

What the best fit says

A least-squares fit through all twenty chips gives a doubling about every 2.1 years. Moore’s 1975 forecast was close to right for half a century. Set the slider to 2.0 and the line tracks the dots closely from the 4004 to the mid-2000s.

The fit hides a bend, though. Fitting only the chips before 2005 gives almost exactly 2.0 years. Fitting only the chips from 2005 on gives about 2.7 years. With ten chips from four makers, that second number is rough, but every chip since 2020 sits below the 2-year line.

Where the trend bent

Moore’s Law counts transistors. For thirty years it came with a bonus: Dennard scaling. Shrink a transistor and it switches faster and uses less power, so power per square millimetre stays flat. Mark Bohr of Intel describes how that broke. By the 65 nm generation the gate oxide was about five atoms thick, and leakage made it hard to lower voltage any further.

Without lower voltage, faster clocks meant more heat. Herb Sutter pointed out in 2005 that clock speeds had flattened around 2003 while transistor counts kept climbing (The Free Lunch Is Over). In 2004 Intel cancelled Tejas, its next single-core Pentium 4, and moved to dual-core designs (EE Times). Sutter predicted that from then on, extra transistors would go into more cores and more cache rather than higher clock speeds.

Cost bent too. Cheaper transistors used to come free with each shrink. In 2014 Zvi Or-Bach of MonolithIC 3D argued that cost per transistor stopped falling at 28 nm. Milind Shah of Google backed the claim at IEDM 2023 (Semiconductor Digest).

The biggest recent dots also count something new: packages, not dies. EPYC Rome’s 39.5 billion transistors sit on nine dies (HEXUS). IBM’s Telum II is a single 600 mm² die and holds 43 billion (IBM Research). Part of the recent growth comes from putting more dies in one package, not from denser dies alone.

So, has it held up?

As a statement about transistor counts, mostly yes. The best fit over fifty years is within about a month of Moore’s 1975 forecast. As a promise of faster, cooler and cheaper chips every two years, no. That promise ended in the mid-2000s. What survives is a count that keeps doubling, now through new transistor shapes, extreme-ultraviolet lithography and many dies in one package, at a slower pace and a higher price.

A note on the data: I list only CPUs, not GPUs or AI accelerators, and I date each chip by the year it shipped. Every value, with its source and any disagreement between sources, is in the “Sources and data” table under the chart.