Moore’s law is the empirical observation that the number of transistors on an integrated circuit doubles about every two years, at a roughly constant manufacturing cost per chip. In semiconductors, it has served for more than half a century as the industry’s planning target for process development, chip design and computer hardware roadmaps.
Moore’s law is not a law of physics. It is an economic and engineering trend that held because manufacturers, equipment suppliers and designers coordinated their investments to keep it true. Its pace has slowed since the mid-2010s, and today the term describes both a historical record and an ongoing debate about how long transistor scaling can continue.
Origin and precise formulation of Moore’s law
Moore’s law originated in a 1965 article by Gordon Moore, then at Fairchild Semiconductor, who noted that the number of components per chip at the lowest cost per component had doubled every year. In 1975 he revised the forecast to a doubling every two years, the version most engineers cite today. The often-quoted 18-month figure came later from industry commentary and referred to performance, not transistor count.
The original formulation was about economics as much as density. Moore observed that each process generation had an optimal level of integration: too few components wasted the chip, while too many reduced yield. The table below shows how the main formulations differ.
| Version | Period | Claim | Status |
|---|---|---|---|
| Original 1965 forecast | 1965-1975 | Components per chip double every year | Held for about a decade |
| Revised 1975 forecast | 1975-2010s | Transistor count doubles every two years | Broadly held for decades |
| Performance variant | 1980s-2000s | Performance doubles about every 18 months | Ended with clock scaling |
| Cost per transistor view | 1965-present | Each transistor becomes cheaper per node | Weakening at leading nodes |
How transistor scaling sustained the trend
Moore’s law was sustained mainly by shrinking transistor dimensions through better photolithography, materials and device structures. Each new process node aimed to reduce linear feature size by about 0.7 times, which roughly halves the area of a given circuit and doubles the transistors that fit on the same silicon.
From the 1970s to the mid-2000s, Dennard scaling made this shrinking even more valuable. It predicted that when transistors shrink, voltage and current scale down as well, so power density stays constant while speed rises. Smaller chips were therefore faster and not hotter, which allowed clock frequencies to climb from a few megahertz to several gigahertz.
Around 2005, Dennard scaling broke down because leakage current and threshold voltage limits stopped supply voltage from falling further. Clock speeds plateaued near 3-5 GHz, and the extra transistors went into multicore processors, larger caches and specialized accelerators instead. Engineers then kept density rising through structural innovations listed below.
| Technology | Introduced around | Problem it addressed |
|---|---|---|
| Strained silicon | 2003 | Low carrier mobility in small channels |
| High-k metal gate | 2007 | Gate oxide leakage at thin dimensions |
| FinFET transistors | 2011 | Poor gate control and short-channel effects |
| EUV lithography | 2019 | Costly multiple patterning with deep ultraviolet |
| Gate-all-around nanosheets | 2022-2025 | Remaining leakage in fin structures |
Moore’s law and related scaling concepts
Moore’s law is often confused with neighboring scaling laws that describe power, cost or performance rather than transistor count. Distinguishing them helps explain why chips still gain transistors while single-thread speed and price behave differently.
| Concept | What it describes | Current relevance |
|---|---|---|
| Moore’s law | Transistor count per chip over time | Continuing at a slower pace |
| Dennard scaling | Constant power density as transistors shrink | Effectively ended around 2005 |
| Rock’s law | Cost of a new fab doubles about every four years | Fabs now cost tens of billions of dollars |
| Koomey’s law | Computations per joule improve over time | Improvement has slowed since 2000 |
| Process node naming | Marketing label for a manufacturing generation | No longer equals a physical dimension |
The last row is a frequent source of confusion. Names such as 5 nm or 3 nm no longer correspond to any single feature on the chip; they indicate a generation with better density, power and performance. Comparing transistor density per square millimeter across foundries gives a more meaningful picture than node names.
Is Moore’s law ending
Whether Moore’s law is ending depends on which formulation is measured. Transistor counts in the largest chips still grow, with leading processors and accelerators reaching tens to over one hundred billion transistors, but the time between density doublings has stretched to roughly two and a half to three years at the leading edge.
The economic version is under more pressure. Lithography tools, design costs and mask sets have become so expensive that the cost per transistor falls only slightly, or not at all, at the newest nodes. Only a handful of semiconductor foundries can still afford leading-edge manufacturing, which concentrates the industry.
Physical limits add further constraints. Gate lengths are approaching a few nanometers, where quantum tunneling, atomic-scale variability and heat removal become hard problems. Experts disagree on timing, and claims that Moore’s law is already dead or fully alive are both oversimplifications.
How the industry continues beyond classic scaling
The industry now pursues More than Moore and system-level scaling, which increase useful transistors per product without relying only on smaller features. These approaches change how designers count progress, moving from transistors per die to transistors per package and performance per watt.
- Chiplets split a large design into smaller dies, improving yield and allowing each part to use the most suitable process node.
- Advanced packaging such as 2.5D interposers and 3D stacking places dies side by side or on top of each other with dense interconnects.
- Domain-specific accelerators, including ASICs and GPUs, raise performance per watt for tasks such as machine learning.
- Backside power delivery moves power wiring under the transistors, freeing routing space on the front side.
- New materials such as two-dimensional semiconductors are under research for future channels.
These methods extend the spirit of Moore’s law, steady improvement in computing capability per unit cost, even where the strict density doubling slows.
Stripped to the essentials
Moore’s law says that engineers manage to fit about twice as many tiny switches, called transistors, onto a computer chip every two years. More switches mean more computing power in the same small space, which is why phones today outperform room-sized computers of the past.
It is a long-running trend, not a rule of nature. It is slowing because transistors are now only a few dozen atoms wide, so the industry also stacks and combines chips to keep progress going.
Where it is applied
A chip design team planning a data center processor for release four years ahead uses Moore’s law as a baseline for its roadmap. Starting from a current design of about 40 billion transistors, the team assumes one or two new process nodes will be available, each offering roughly 1.6 to 1.8 times the logic density rather than the classic doubling.
Because a single die would exceed practical size and yield limits, the architects split the product into several chiplets: compute dies on the newest node and input-output dies on an older, cheaper node, joined with advanced packaging. The resulting package holds well over 100 billion transistors, and the team budgets power carefully, since the end of Dennard scaling means more transistors no longer come with free energy savings.