tequila bomb

Intel’s first mobile silicon comes of age… Or not.

Jon Stokes does a great job of giving us an overview of Moorestown. Intel’s first x86-based SoC aimed at the smartphone market. Which will from now on be known officially as the Atom Z6XX series. He quotes benchmarking figures from AnandTech, which in the most part, came from Intel.

The conclusion I draw from it all is that Moorestown is a cutdown desktop chip, paired with an over-clocked mobile GPU, and all crammed rather haphazardly into a small hamlet of silicon. And it sucks power if you use it for anything other than making calls.

If I can steal from Stacey Higginbotham over at GIGAOM :

ARM has always had an advantage in mobile because the chips based on the instruction set were designed to sip power rather than glug it. That translates into a longer battery life and presumably a smaller form factor for the battery and end device. The biggest hurdle it had was that most of the software people want to use is designed to run on x86 chips. But in the last two years, thanks to the efforts to port software to its instruction set, and the overall movement of applications to the web, ARM has whittled away Intel’s advantage on that front, and we’re now seeing ARM-based devices finally hit the consumer market. Faced with faster processors, longer battery life and always-on connectivity, I agree with ABI that ARM will blow Intel’s Atom based gadgets out of the water.

In my opinion ARM have nothing to worry about this generation of mobile silicon from Intel. Or for quite some time.

Here are the highlights from Jon’s article :

At the heart of Moorestown is the application processor, which is an SoC (system on a chip) that features an in-order, Atom-derived x86 CPU core and a PowerVR SGX GPU from Imagination technologies.

…

When you include the separate memory chip, which competitors like Apple’s A4 have integrated into the SoC package, you’re looking at five chips in all for the Moorestown platform. Compare this to Qualcomm’s Snapdragon, which integrates most of this functionality, including the baseband processor, into a single package. Moorestown’s sprawling five-chip layout will make it less efficient and more expensive than a comparable, more highly integrated ARM solution.

…

As for performance, this is where Intel really makes some claims for Moorestown. Anand reports that he saw Quake 3 running on Moorestown at 100FPS. He also reports that Moorestown completed the SunSpider benchmark in a scorching two seconds (compare about 10 seconds for a 1GHz Snapdragon).

The GPU in Moorsetown has been clocked at around twice the speed it would normally be in an ARM SoC. And this is where the problems start…

If you compare Intel’s power consumption numbers to its performance numbers, you’ll notice that there’s zero overlap. Nowhere does Intel reveal how much power Moorestown draws when running Quake 3 at 100FPS, or when executing the SunSpider JavaScript benchmark in two seconds. That’s because the power draw in those situations is certain to be astronomic by smartphone standards.

This is standard fare from Intel, who are quite happy to use slight of hand to obfuscate performance figures they don’t want people to see. They have done far worse when hiding IGP problems.

The reason for Moorestown’s decidedly unsmartphone-like range of power/performance points is that Intel relied heavily on power gating and other, very fine-grained dynamic power optimization techniques.

The level of OS involvement needed to manage this very thorough and carefully implemented level of dynamic power management is substantial.

In this respect, Moorestown is a tiny bit like the late Larrabee project, because Moorestown’s substantial software component will be critical to making the platform competitive in its initial incarnation.

The massive problem there is that Moorsetown can’t be let out in the wild as application silicon for any manufacturer to use. It needs a tightly coordinated OS support system around it. Which is why currently it’s only being used by one vendor consortium, who are partnered with Intel directly : MeeGo

You won’t just fire up Ubuntu—or even Chrome OS or Android or an x86 port of webOS—on your Moorestown-based tablet and expect to get any kind of battery life from it.

Jon’s conclusion is that…

Intel’s first serious smartphone part will be the 32nm Medfield in 2011. Medfield will combine the application processor and PCH onto a single chip, and I wouldn’t be surprised to see it sport in-package memory. In all, Medfield will be a real ARM competitor—and then it’ll be game on.

I am not so convinced. Intel has failed with Larabee. Its Core series IGPs are still not competitive. And now it is producing mobile-silicon which is power hungry and heavilly restricted in terms of what OS it can run because of specialist power management needs. In fact Moorsetown is almost a closed system. And it still has a bloated desktop instruction set.

Intel should buy ARM, stop cutting Nvidia out of the GPU loop, and concentrate on what it is good at. Desktop CPUs.

New Intel® Atom™ Press Release.