dark silicon
Imagine a house wired with a hundred light bulbs, but the fuse box can only safely carry enough current for thirty of them at once. You can install all hundred, and choose which thirty to switch on at any moment, but you can never light them all together without blowing the fuse. Modern chips are in exactly this bind. 'Dark silicon' is the large fraction of a chip's transistors that, at any given instant, must stay dark — unpowered or idle — because powering them all at full speed would exceed the power and thermal budget.
Here is how it arose. Moore's law kept doubling the transistors that fit on a chip, but the end of Dennard scaling meant the power per transistor stopped falling at the same pace. So each new generation packs many more transistors than it can afford to switch simultaneously within a fixed power-and-heat budget. The arithmetic is stark: a chip might have, say, four times the transistors of an earlier one but only a small increase in its power ceiling, so a growing share — sometimes a majority at advanced nodes — cannot be active at once. That permanently-or-temporarily-off share is the dark silicon.
Dark silicon is not pure waste; it reshaped how chips are designed, and that is the honest, important part. If you cannot light everything at once, then spend the spare transistors on many specialized blocks — a video decoder, an encryption engine, an AI accelerator, several CPU and GPU cores — and light up only the few best-suited to the current task, leaving the rest dark. This is a major driver of heterogeneous, specialized systems-on-chip: the constraint turned 'do everything with one fast core' into 'keep a toolbox of accelerators, most of them dark most of the time.' Dark silicon, the power wall, and the end of Dennard scaling are three faces of the same underlying limit that ended the simple scaling era.
A phone SoC contains CPU cores, a GPU, an image processor, a video codec block, and a neural accelerator. While you shoot video, the camera and codec blocks light up and the neural accelerator may be idle (dark); when you run face unlock, the neural block lights up and the codec goes dark. The chip is never fully lit at once — that is dark silicon working as intended.
Dark silicon turns a constraint into a strategy: many specialized blocks, only the useful few lit at any moment.
Dark silicon is a consequence of Moore's law continuing while Dennard scaling ended — transistor counts grew but power-per-transistor did not fall enough. It is why specialization and heterogeneity, not ever-bigger uniform cores, became the path forward.