Is Koomey's Law eclipsing Moore's law? | Hacker News Reader

Is Koomey's Law eclipsing Moore's law?

economicsofinformation.com

28 pointsmikecane12 comments

Exponential increase of transistor power efficiency was predicted by Dennard's scaling laws, so it sounds like Koomey has just empirically verified it. It's an interesting time to publish such a study, though, considering that everyone is discussing the end of Dennard scaling.

http://www.ieee.org/portal/cms_docs_societies/sscs/PrintEdit...

A link to the real article from MIT's Technology Review: http://www.technologyreview.com/printer_friendly_article.asp...

Why does everyone get Moore's law wrong? He said: "The number of transistors that can be placed inexpensively on an integrated circuit doubles approximately every two years". I guess that "Computing Power" and transistors are related, but he was really just talking about more transistors on the chip.

Are these two laws even mutually exclusive?

Only if you assume that there's a one-to-one correspondence between transistors and processing power. That is, it doesn't take into account improved microarchitecture (e.g. speculative execution, pipelines and cache management) that more efficiently use a given number of transistors.

Isn't the different architecture more of an "if you want to meaningfully use more transistors, you will have to do it something differently" than of an "we can do faster if we do things differently?"

I would expect that any micro-architecture effects would be relatively marginal, relative to the rough number Moore's law gives.

So, I think, at the resolution of Moore's law, "number of transistors double"' and "processing power doubles" are equivalent.

Another related 'law' is "CPUs gain a bit every 18 months". That metric predicts 6 years between outgrowing 4 bits and outgrowing 8bit CPUs, 12 years before needing 16 bits, 24 for needing 64 bits, and indicates we will not need a 128bits CPU on the desktop before 2050 or so (of course, all if we manage to stretch Moore's law that far)

I think that is in the right ballpark.

No, but decreasing power consumption may more important than increasing computing power for mobile devices.

Gordon Moore never defined anything he called "Moore's Law". Rather he started talking about things doublng periodically - first transistors/device but later other sorts of things too - and after a decade people started to refer to any CPU related periodic doubling as "Moore's Law". Moore even wrote an memo endorsing this general usage.

More transistors on a die is frequently a result of an improved process. Moving to a smaller process means you can run your clock at a higher rate (generally speaking), which often (not always) translates to faster operations.

There is a striking parallel between this and American cars in the 50's and 60's, when the focus powertrain-wise was simply to get as much power out of an engine at the least possible cost, efficiency be damned.

So mobile devices are sort of the Arab Oil Embargo of computing.

It seems to me like 'Koomey's Law' is just an emergent property of Moore's Law.

Not entirely, do you remember Intel's NetBurst fiasco in early Pentium 4 CPUs? https://secure.wikimedia.org/wikipedia/en/wiki/Netburst#Succ... Power usage is free to increase until you physically have trouble removing heat as fast as you're generating it. Today, you can crank a i7 CPU to 4.0GHz without much trouble, if you have decent cooling. The architecture affects both performance and power consumption. Just having smaller transistors isn't the whole story.

Reply on news.ycombinator.com