Whither Moore's Law; Introducing Koomey's Law

Joining the ranks of accepted submitters, Beorytis writes "MIT Technology review reports on a recent paper by Stanford professor Dr. Jon Koomey, which claims to show that the energy efficiency of computing doubles every 1.5 years. Note that efficiency is considered in terms of a fixed computing load, a point soon to be lost on the mainstream press. Also interesting is a graph in a related blog post that really highlights the meaning of the 'fixed computing load' assumption by plotting computations per kWh vs. time. An early hobbyist computer, the Altair 8800 sits right near the Cray-1 supercomputer of the same era."

Comments

Power Hog

[Waffle Iron]

My favorite example of computing (in)efficiency is the USAF's SAGE bomber tracking computers introduced in the 1950s. These vacuum tube machines had CPU horsepower probably in the same ballpark as an 80286, but could draw more than 2 megawatts of power each. They didn't decommission the last one until the 1980s.

Re:Power Hog

[anubi]

Even the idea one could even implement a vacuum-tube machine capable of performing at 286-levels to me is a miracle in itself. 6502 maybe, but, to me, even the lowly 286 represents a level of sophistication I could not even imagine being implemented with vacuum-tube technology.

I've never seen a SAGE, but it must have been quite a machine. In my imagination, it must have been about the size of a Wal-Mart. With the physical size of the thing, it would amaze me that they would be able to clock the thing anything more than 100 KHz or so.

Yes, I do know what a 6SN7 is. And a 12AT7, which I suspect the machine was full of ( or its JAN equivalent).

Do the designations 12SA7, 12SK7, 12SQ7, 50L6, 35Z5 still ring a bell with anyone?

Re:Theoretical limits?

[PaulBu]

Yes, there is if you "erase" intermediate results -- look up 'von Neumann-Landauer limit', kT*ln(2) energy must be dissipated for non-reversible computation.

Reversible computation can theoretically approach zero energy dissipation.

Wikipedia is your friend! :)

Paul B.

Re:Theoretical limits?

[MajroMax]

Without reversible computing, there indeed is a fundamental limit to how much energy a computation takes. In short, "erasing" one bit of data adds entropy to a system, so it must dissipate kT ln 2 energy to heat. This is an extremely odd intersection between the information theoretic notion of entropy and the physical notion of entropy.

Since the energy is only required when information is erased, reversible computing can get around this requirement. Aside from basic physics-level problems with building these logic gates, the problem with reversible computing is that it effectively requires keeping each intermediate result. Still, once we get down to anywhere close to the kT ln 2 physical constraint, reversible logic is going to look very attractive.

True, but...

[PaulBu]

I do not think that you get net energy savings (by using the same basic technology, e.g., CMOS at room temeprature or "cold"), if you take into account the fact that cooling things down also costs energy! For example, liquid helium refrigeration costs about 1 kW of wall outlet power to compensate for 1 W dissipated at 4.2 K.

Changing your basic technology to, e.g., some version of superconductor-based logic can help (a lot!), current state of the art (in my very biased opinion, since I am cheering for those guys, and have been involved in related research for years) is here: http://spectrum.ieee.org/semiconductors/design/superconductor-logic-goes-lowpower ...

Paul B.

Re:Theoretical limits?

[bunratty]

Yes, reversible computation can theoretically approach zero energy dissipation, but if you use no energy, the computation is just as likely to run forwards as backwards. You still need to consume energy to get the computation to make progress in one direction or the other. Richard Feynman has a good description of this idea in his Lectures on Computation.

This is such an absurd point

[terraformer]

It's the inverse of Moore's law so yeah, duh....

If your compute power doubles in the same size die every 1.5 years, then if you halve the die size keeping the compute power the same you actually cut the power in half. This is a very well known phenomenon and Koomey is doing what he has been for a while, making headlines with little substance and lots of flair.

That Microsoft and Intel paid for this research calls into question what it was they were actually paying for.