Dennard scaling
/ DEN-ard SKAY-ling /
Moore's law gave us more transistors; Dennard scaling was the happy companion that, for decades, turned those extra transistors into faster, not hotter, chips. Named after Robert Dennard, it observed that as transistors shrank, you could also lower their voltage in step, so that the power used per unit of chip area stayed roughly constant even as the switches got smaller, faster, and far more numerous. In short: you got more and faster transistors essentially for free, with no rise in heat per area.
Concretely, power in a chip roughly follows voltage squared times frequency times capacitance. As Dennard scaling shrank transistors and dropped voltage together, designers could pack more transistors and clock them faster while the total power density held steady. This is the real engine behind the era when clock rates soared and every new chip simply ran faster than the last — the comfortable years when waiting for next year's computer was a sound plan.
Here is the pivotal, honest fact: Dennard scaling effectively ended around 2005. Voltage could no longer keep dropping, because at tiny sizes transistors leak current even when off (static or leakage power), so shrinking further no longer keeps power density flat — it rises. The consequence is the power wall: you cannot just keep cranking the clock without melting the chip. This breakdown, not Moore's law alone, is why the industry pivoted to multicore processors, specialised accelerators, and energy-aware design. The end of the free lunch is the reason architecture became urgent again.
Before about 2005, each shrink let voltage drop too, so a faster, denser chip used about the same power per area. After it ended, leakage made power density climb — so doubling cores at a fixed clock replaced simply racing the clock upward.
When Dennard scaling broke, the free clock-speed ride ended and multicore began.
Do not confuse the two laws: Moore's law is about transistor count, Dennard scaling about power-per-area as transistors shrink. Moore's slowed; Dennard's effectively ended around 2005 — and that ending caused the multicore turn.