OSNews

This comparison is horribly flawed. The most prominent error is that both the TDP number you quoted and PA-SEMI numbers are with both cores running, SPECint only uses one core, this alone would make this comparison moot but still there are other problems:

The comparison is valid because it establishes a lower bound on the SPECint performance/watt of each CPU. In general, both CPUs will achieve higher ratios, because SPECint will only cause the CPU to reach some fraction of its maximum TDP. This is both because it exercises only one core, and because it won't use the FP hardware at all. For a ballpark estimation like this, we can assume the fraction to be similar on each processor, which would cause it to factor out, leaving the relative performance/watt similar regardless of its value.

First of all Intel TDP are much closer to IBM et al typical power numbers whilst the 25W figure from PA-SEMI slides is under a thermal virus, the typical power quoted was 17W.

There is a nice thread on RWT this week that talks about how Intel measures their TDP under a power virus, and the measured power usage under such code is actually less than the stated TDP.

Now the 17W figure also includes power consumed by the I/O stuff, so please put a part of the northbridge+soutbridge power consumption into your C2D figure.

That's a fair point. Since we have actual per-core figures for the PA6T, let's use that instead. So the 2.0 GHz 1682M would achieve 1200 SPECint at 14W, giving it a ratio of 86. The 1.08 GHz Core 2 ULV would achieve 1150 SPECint at 9W, giving it a ratio of 127. Better, but still not competitive with what Intel has now, much less what they'll have in Q4 when this thing is actually shipping in volume.

Last but not least Intel's ULV parts are cherry picked out of the best binsplits of a certain process node whilst 1682M hits that power levels even under significant process variability. Vdd can be adjusted to trade between static and dynamic power consumption over the production to increase yield while keeping the processor in its power envelope.

While that's a fair point, Intel also has vastly more production resources to cherry-pick from than whomever PA-Semi outsources their fabrication to. If we're talking about right now, I'm sure Intel has a lot more ULV Core 2's than PA-Semi has limited-production 1682M's to sell. If we're talking about Q4, when the 1682M will be in volume production, well, then Intel will be shipping 45nm CPUs, and won't have to cherry-pick.

Finally C2D is being produced on the most bleeding edge process by Intel whilst the 1682M will be fabbed, Intel process is specifically tailored for high performance MPUs.

Yep. But high-performance process capability is a major competitive advantage for Intel. Apple doesn't care if the design would be competitive if fabbed on Intel's process, because that's not a realistic design proposition.

The 1682M's performance/watt looks good if you actually leverage it's extra features (fast FP, integrated I/O). However, if you just look at integer performance/watt, it's not that great compared to Intel, and the absolute performance even at 2 GHz is really low. Now, obviously there are a huge number of customers for which this chip's performance characteristics are absolutely right. No doubt PA-Semi intends to sell to those people. However, pointing to this chip as refutation of the reasoning behind the switch is silly. It's a validation of that reasoning: it's yet another embedded chip that's not designed for Apple's target market. It's not easy to design high-performance CPUs for the desktop market, and if you're not even trying, but instead repurposing chips designed for different markets, you're never going to be competitive!