Moore's Law Doubling

In April 1965, Gordon Moore was asked to write a short piece for the thirty-fifth anniversary issue of Electronics. He had four or five data points. He plotted, for each year, the number of components on the integrated circuit with the lowest cost per component, noticed the figure had been doubling annually, and projected the trend a decade forward to about 65,000 components by 1975. He was roughly right. In 1975 he revised the rate to a doubling every two years, and that is the figure the industry ran on for the next thirty.

The argument in the paper was economic, not physical. At any moment there is a chip complexity that minimises cost per component: below it you are wasting the fixed costs of packaging and testing, above it yield losses dominate. Moore's observation was that this optimum was moving, and moving exponentially. Nothing in the article says physics permits a doubling every two years — it says the cost curve's minimum has been sliding at that rate. That distinction is the single most misunderstood thing about the whole subject.

What made it hold for four decades was that it became a coordination device. The industry adopted it as a planning target; the International Technology Roadmap for Semiconductors codified it; equipment makers, materials suppliers and fabs all committed capital on the assumption that everyone else would meet the schedule too. A prophecy that organises the investment required to fulfil it is not a law of nature, and calling it one gets the causation exactly backwards.

Where it stands now depends on what you count. Transistors per chip still rise, though the doubling period has stretched past two years and increasingly comes from stacking and advanced packaging rather than lithography alone. Cost per transistor — Moore's actual subject — stopped falling reliably somewhere around the 28 nm node, which is why so many designs stayed there for so long. Dennard scaling, the separate curve that made each new node cooler as well as denser, ended around 2005 and took the free performance with it. And the node names stopped meaning anything physical years ago: "5 nm" names a process generation, not a dimension you could measure on the die. When someone says Moore's law is dead or alive, ask which of those they mean.