Hacker News Reader

Comment by monocasa | Hacker News Reader

This is the Dennard of Dennard Scaling, a chip scaling law that is arguably as important as Moore's Law.

https://en.wikipedia.org/wiki/Dennard_scaling

The end of Dennard scaling was why the Pentium 4 architecture was a dead end and never hit 10Ghz like it was supposed to, why the Cell processor never hit the 5Ghz it was supposed to, why we've been spending quite a bit of the transistor budget on more cores rather than a very fancy single CPU core of 10Bs of transistors, and why chips with lower thermal limits will see a lot of "dead silicon" where you can't actually light up the whole chip at once without melting it.

View on HN

Dennard scaling, for people in the industry, was far more important than Moore's law when it was available.

Moore made a high-level observation, but Dennard told you how to do it.

>The end of Dennard scaling was why the Pentium 4 architecture was a dead end and never hit 10Ghz like it was supposed to, why the Cell processor never hit the 5Ghz it was supposed to

Around that time the PowerPC 970 aka G5 also failed to achieve 3 GHz, breaking the promise Steve Jobs publicly made at one of his keynotes for Apple.

But when Dennard Scaling was in full swing, it was glorious. It only took six years to go from the original Pentium 60Hz on March 22, 1993 to Pentium III 600 MHz on August 2, 1999 and just a year later you could buy a 1GHz one.

> The end of Dennard scaling was why the Pentium 4 architecture was a dead end and never hit 10Ghz like it was supposed to

I've always been puzzled by this. Did Intel really not see this coming? I remember talking to Intel engineers way back when they were promising 10GHz in the near future - I think the codename at the time was Tejas. They seemed very confident. The architecture must have already been planned out - and yet it seems from the outside like the end of Dennard scaling was a total surprise to them?

Intel (and almost everyone else tbh) didn’t fully appreciate how Denard scaling would play out at smaller nodes. They expected to keep lowering the transistor threshold voltage alongside transistor size but that became increasingly difficult due to leakage currents.

They also played with tricks like strained silicon on 90nm and high-k metal gate in 45 nm in order to boost performance and lower leakage respectively.

It was a scaling law that had worked for three decades, and didn't show any signs of faltering. Even the most senior of the people building those designs had spent their entire careers in a world where it just was true.

And then it went away in an instant.

Now there will be a twinge of sadness whenever I read a paper beginning with “Since the end of Dennard scaling…”.

"The end of Dennard scaling and the impending repeal of Moore's law" is very overused in very many papers. :)

Well, it's basically the technical implementation of Moore's law, since Moore's law is just an empirical observation. (And maybe also a self-fulfilling prophecy)

Reply on news.ycombinator.com