Moore's Law

Moore’s Law is the shorthand people use when they argue about whether computers keep getting cheaper at a steady clip. The part that mattered for software was never the name. It was the long window where lithography, yield, and design work lined up so teams could count on more transistors per dollar on a drumbeat they could plan against.

In 1975, I updated it, adding the purple line, and I argued why the slope was going to change from doubling every year to doubling every two years.

That line is Gordon Moore in 2003, looking back at his 1975 revision, in Intel’s archived ISSCC keynote transcript. It is the cadence most speakers mean when they say “Moore’s Law” in a deck.

The same source shows why the curve felt like free money for programs. Transistor shipments exploded while average transistor cost fell off a cliff. Whole generations of systems could trade human time for brute silicon on throughput, RAM headroom, single-node parallelism, and waste the compiler never cleaned up.

Moore also warns in the same keynote that “no physical quantity can continue to change exponentially forever.” We live in the bend now. The Wikipedia article on Moore’s law summarizes field reporting that cadence slipped industry-wide below the textbook two-year rhythm after about 2010, with Intel executives at times describing effective doublings on the order of two-and-a-half to three years instead of two.

What that means for builders is direct.

• You cannot assume the next node erases bad asymptotics or sloppy constant factors. Amdahl’s law still caps you on serial work no matter how many cores you buy.
• You pay for wins with architecture choices (domain-specific accelerators, tighter data layouts, fewer round trips) and with coordination cost when you spread work out.
• Power and cooling sit next to dollars on the same spreadsheet because the old trick of shrinking features while holding power density flat stopped working in the way Dennard scaling promised.

The bill for fabs and lithography went the other way the whole time (Moore’s second law on the same page, under that heading). Chip cost per transistor fell for buyers while capital intensity rose for foundries. Software rode the buyer curve. Planning still has to respect the producer curve when you pick vendors, contract for capacity, or model how fast bespoke hardware can refresh.

The honest posture is to treat cadence like any other dependency. Bank on physics, chemistry, and capital investment the way you bank on a critical vendor, with monitoring and a Plan B when the slope changes.