Dennard scaling
Also known as: Dennard's law, MOSFET scaling, end of Dennard scaling
Dennard scaling is the observation that as transistors shrink, their power density stays roughly constant, so smaller transistors can run faster without making chips hotter.
It powered decades of rising clock speeds alongside Moore's Law. While Moore's Law gets the fame, Dennard scaling's breakdown in the mid-2000s is the real reason your computer stopped getting dramatically faster in clock speed and started adding more cores instead.
Dennard scaling, explained at three levels
At home
Imagine shrinking a car to half its size, engine included, so it burns proportionally less fuel and can rev faster without overheating. For about 30 years, transistors behaved that way: each time they shrank, they used proportionally less power, so chipmakers could pack in more and run them faster without the chip getting hotter. That rule is Dennard scaling. Around 2005, it stopped working.
At work
Dennard scaling is a 1974 rule from IBM engineer Robert Dennard: as transistors shrink, their power use falls in step, so power per square millimetre stays constant. It let processors gain speed every year without running hotter. It broke down around 2005. For businesses, that's why new computers no longer make old single-threaded software dramatically faster, and why performance now comes from more cores and specialized chips.
At home
Dennard showed that if you shrink a transistor's dimensions and its operating voltage by the same factor, it switches faster and uses less power, while power per area stays the same. That's what let clock speeds climb from megahertz to gigahertz. The plan depended on voltage falling with every generation. Once voltage got stuck near one volt, smaller, faster transistors no longer used proportionally less power, so heat rose and clock speeds stalled at a few gigahertz.
At work
When Dennard scaling ended, chipmakers kept getting more transistors from Moore's Law but couldn't run them all faster. They spent them differently: multiple cores instead of one faster core, specialized accelerators such as GPUs and NPUs, and larger caches. That shift moved the burden of speed onto software. A program that runs on one thread gets little benefit from new hardware, while code written to run in parallel scales with each generation. That affects how teams plan application modernization.
At home
A chip's switching power is roughly capacitance times voltage squared times frequency. Shrinking transistors reduces capacitance, and lowering voltage cuts power sharply because it's squared. But a transistor needs a minimum threshold voltage to switch off cleanly, and lowering that threshold makes off-state leakage grow exponentially. So supply voltage stalled near one volt, and power density began rising. While it sounds like a narrow physics detail, in my opinion that single limit reshaped the industry, pushing it toward multicore and specialized chips.
At work
Post-Dennard design is about performance per watt. Techniques include dynamic voltage and frequency scaling, which adjusts speed to load; power gating, which shuts off idle blocks; and heterogeneous cores, such as Intel's performance and efficiency cores or Arm's big.LITTLE design. Because not every transistor can run at full power at once, some parts of a chip must stay idle, an effect called dark silicon. Data centre economics follow the same logic, with energy and cooling now central to compute costs.
Real-world examples
At home
Marcus wondered why his 10-year-old laptop's processor ran at a similar gigahertz speed to his new one. He learned that clock speeds stopped rising after Dennard scaling ended, and new chips got faster mainly through more cores and better design.
At work
Fatima is an IT architect planning to move an old single-threaded application to new servers. She realized new hardware wouldn't speed it up much because of the limits after Dennard scaling. Her team redesigned the slowest part to run in parallel instead.
How it compares
| feature | Dennard scaling | Moore's Law |
|---|---|---|
| What it describes | Power density staying constant as transistors shrink | Transistor counts doubling roughly every two years |
| Status | Broke down around 2005 | Slowing but continuing in modified forms |
| Main effect | Rising clock speeds | More transistors per chip |
| Who it's named for | Robert Dennard (IBM) | Gordon Moore (Intel co-founder) |
Moore's Law gave us more transistors; Dennard scaling let them run faster. Only the first lasted.
For analysts
Don't assume hardware upgrades will fix slow software. I'd profile applications before buying faster machines, since the mid-2000s many performance problems are software design problems.
History & origin
Robert Dennard, the IBM engineer who invented DRAM memory in the 1960s, co-authored a landmark 1974 paper describing how scaling down transistor dimensions, voltages, and currents together would keep power density constant. For about three decades, it held: processors went from megahertz to gigahertz. Around 2005, it broke down because voltages couldn't drop further without leakage current exploding. Intel cancelled its planned high-frequency Tejas processor in 2004 and turned to multicore chips, and the industry followed. The concept of 'dark silicon,' where parts of a chip must stay off to avoid overheating, emerged in the early 2010s. In my view, the end of Dennard scaling is one of the most important and least appreciated turning points in computing history.