Dennard Scaling

Dennard scaling is the scaling law, formulated in a 1974 paper led by IBM's Robert Dennard, stating that as transistors shrink, their power density stays roughly constant. In its ideal form: scale a transistor's linear dimensions down by some factor, scale its operating voltage and current down by the same factor, and power per unit of chip area holds steady — even as you pack in more transistors that switch faster. For roughly three decades this was the quiet engine underneath the entire semiconductor industry, and it is the deep reason chips once got denser, faster, and more efficient all at once, with no trade-off demanded.

The free lunch it provided

From the mid-1970s to the mid-2000s, each new process node delivered a compounding miracle: more transistors, higher clock speeds, and a flat power budget. Moore's Law gets the popular credit, but Moore's Law only promises more transistors — Dennard scaling is what made those transistors affordable to run. Software got faster without being rewritten; laptops got cooler while getting quicker; and computation per joule improved on a schedule you could plan a business around. Had Bitcoin existed in that era, mining efficiency would have improved almost automatically with every shrink, no cleverness required — a J/TH curve descending on fab-industry autopilot.

The breakdown around 2005

The law died of leakage. Voltage scaling required lowering transistor threshold voltages in step with size, but subthreshold conduction — current sneaking through a transistor that is nominally off — rises exponentially as threshold voltage falls, a limit rooted in the roughly 60 mV/decade subthreshold-swing floor of conventional transistors at room temperature. By 2005–2006, supply voltages could no longer fall meaningfully, so each shrink now packed more switching activity into the same area at nearly the same voltage: power density began climbing instead of holding flat. Clock speeds stalled in the few-gigahertz range, where they largely remain, and the industry pivoted — to multi-core processors (more parallel units instead of faster ones), to 3D transistor geometries like the FinFET that restored electrostatic control and cut leakage, and to aggressive power-management techniques such as clock gating and power gating that fight leakage current block by block. Note the distinction worth keeping sharp: Moore's Law (transistor count) slowed but continued; Dennard scaling (free efficiency) ended. The industry entered the era of dark silicon — chips with more transistors than their power budget can afford to run simultaneously.

Why mining lives in the post-Dennard world

Every Bitcoin mining ASIC is engineered against the wreckage of this law. Efficiency gains no longer fall out of the process node for free — they are clawed out with domain-level voltage tuning (always per domain, never per chip), runtime autotuning that finds each chip's real voltage-frequency floor, ruthless leakage management, and thermal design that treats every joule as the enemy it is. The gap between generations tells the story: the efficiency ladder from ~98 J/TH silicon in 2016-era machines to modern sub-20 J/TH parts owes as much to architecture and power engineering as to lithography. Dennard's law ending is also, indirectly, why proof-of-work works as a security model: with no free efficiency left, hashrate must be purchased with real capital and real energy — and no one, anywhere, gets a shortcut.

The law's afterlife is also a fair warning against extrapolation: the semiconductor industry's most reliable rule held for thirty years and then stopped, not because anyone repealed it but because physics ran out. Plan hardware purchases accordingly — efficiency gains between ASIC generations now arrive in engineered increments, bought with design effort, not delivered by calendar. For the home miner this is quietly good news: a well-tuned machine from the current generation stays competitive longer than the Dennard-era upgrade treadmill ever allowed, and tuning skill compounds while silicon merely depreciates.