AM (myname4rwt@jee.male) on 9/28/09 wrote:
---------------------------
>.. is something I've been wondering about. I mean, face it: to say that power efficiency
>of available Atoms compared to available ULV Core2s is unimpressive is an undeserved compliment.
>
>Power efficiency of Atom's main bins (4W/core, 1.6 GHz, 512k/core)

http://ark.intel.com/ProductCollection.aspx?familyID=29035

Atom 230 isn't a 'main bin'. It's the bin for leaky parts which Intel salvages by selling them for use in desktops rather than the main volume driver for Atom (netbooks).

The Z series Atoms are relatively recent developments, and are now likely the most representative. But if you're looking for the most common Atom bin by number of chips shipped to date, that would be the N270 (2.5W @ 1.60GHz), the backbone of the netbook revolution.

>is clearly worse if you consider there have been 10W DC 1.6 GHz 3M Core2s (SU9600) for some time. And quickly checking for benchmarks vs Core2, e.g. I used this: http://www.tomshardware.com/reviews/intel-atom-efficiency,2069-8.html
>it appears low-power 2.5W bin hardly makes it a winner on the metric it was allegedly
>optimized for. And power-efficiency at platform level -- just forget it.

The THG article you linked to compares the performance of a 65W TDP 2.53GHz C2D to a 4W 1.6GHz Atom 230, not a 10W 1.6GHz SU9600 to a 2.5W 1.6GHz N270 as you imply.

And if you think the power figures measured by THG are meaningful, all they really prove is that chipsets matter. A lot. Atom 230 is normally paired with the 945GC GMCH (22.2W) and the ICH7 I/O hub (3.3W), for a total TDP of 29.5W. By "normally paired", I mean that this is the only Intel-approved chipset for Atom 230 (chipset pairings are shown for most processors at ark.intel.com). Intel puts (pricing?) pressure on Atom system designers to get them to use only these approved combos, and it seems to work because you almost never find anything else. The board reviewed by Tom's was not an exception.

In other words, for reasons known only to Intel, Atom desktop systems are almost always crippled (from a power perspective) by a horribly old (130nm!) and power hungry desktop chipset.

The chipset used in netbooks with the Atom N270 isn't much better: the 945GSE, a 6W TDP mobile version of 945GC. The I/O hub is still ICH7. The total TDP of this combo is 11.8W, much less than half the power of the desktop Atom configuration.

But if you look at what is available for pairing with the more current Z series Atoms, Intel has finally come up with a decent chipset: the US15L, a 2.3W TDP single chip MCH+ICH.

Now consider that the Atom and the US15L still have to communicate by pushing a huge number of FSB signals through pins and PCB traces. How much of the TDP on each side is FSB I/O power which will disappear when Atom SoCs arrive?

If I have a point here, it is this: with a CPU whose TDP is as tiny as Atom's (even the leaky-bin Atom 230), these chipset details are hugely important to nail down. Especially if you're trying to talk sensibly about the power efficiency of the CPU core alone based on measurements of whole-system power. To date, the vast majority of Atoms have been sold paired with an *I/O hub* (the ICH7) which uses more power than the processor itself.

Finally, I would note that there has been a rather large blind spot in much of this discussion: die size. The SU9600 has a die literally 4x bigger than Atom N270. (And 4x the TDP too.) Even after accounting for the fact that SU9600 has 2x as many processor cores, it's still not terribly surprising that 2x the die area and power budget results in a faster CPU core. The flip side: it's economical for Intel to sell the Atom at much, much lower prices.