AMD vs Intel: Computex Round Two, Introducing The Athlon II

When I found out that AMD was shipping four-core Phenom CPUs as three-core devices, I thought the company was having serious 65 nm process woes. And when I more recently discovered that AMD was selling two-of-four-core-disabled chips labeled as high-end Athlon X2s, I didn’t know what to think. Granted, it’s nice to have an architecture that provides for such partially yielding flexibility, much as modern GPUs do, and especially when you’re a small company with a limited number of chip designers at your disposal. But as I wrote a bit over a year ago:

I still get stuck on the thought that AMD’s shipping a tangible amount of Phenom product with a notable amount of per-die silicon area disabled, therefore non-revenue-generating. If you were AMD, wouldn’t you rather ship each die as a higher-revenue quad-core chip?

As such, I’ve long suspected that AMD’s been frantically working behind the scenes on a true single- and/or dual-core (therefore die size-optimized) CPU derived from the same K10 microarchitecture that’s formed the foundation of the last few generations of chips. And the company followed through on my forecast with the just-unveiled Athlon II CPU family. When I first heard of the products a few weeks ago under nondisclosure, I wondered if, based on their names, AMD had simply further migrated its K8 microarchitecture-based Athlon 64 X2 processors (originally introduced four years ago on a 90 nm process, and later litho-shrunk to 65 nm along with dropping the ‘64′ portion of their moniker) down to 45 nm.

As it thankfully turns out, Athlon II is also K10-derived, with a cache twist that I didn’t anticipate upfront but which makes sense upon further analysis. Unlike the company’s quad-core (whether disabled or activated) CPUs, which share a common pool of 6 MBytes (Phenom II, Shanghai and Istanbul) or 2 MBytes (Phenom and Barcelona) of L3 cache, Athlon II completely dispenses with the L3 memory array as a further die-slimming exercise. In exchange, the Athlon II architecture doubles the amount of L2 cache (both as compared to most K8-based CPUs and to all prior K10-based CPUs) from 512 KBytes to 1 MByte per core, or said another way, to 2 MBytes total per-die for Athlon II X2 CPU variants. Strictly speaking, some early Athlon 64 X2s only had 256 KBytes of L2 per core, but I digress…

Here’s a die shot of the dual-core end result:

So far, AMD’s only announced one Athlon II variant, the dual-core model X2 250 running at a 3 GHz core frequency and with an $87 estimated single-unit retail price tag. Its AM3 pinout is compatible with both AM2+ and AM3 sockets, and therefore with both DDR2 and DDR3 SDRAM modules. And the company claims that the chip’s TDP is only 65W, to boot.

Note that AMD was very careful to refer to this product family as Athlon II (no more, no less) while only adding the X2 dual-core clarifier in conjunction with the initial announced model number. While such nebulousness allows, for example, the company to further increment its products’ core counts in the future, it also enables AMD to potentially offer single-core Athlon II variants. Company officials I spoke with were adamant that there are no plans to further litho-shrink the K8-based Athlon architecture below 65 nm. But as much as AMD strives to publicly poo-poo Intel’s Atom (and as much as I think that in Atom Intel’s created a profit-strapped beast over which it no longer has absolute control), at some point AMD is going to have to offer up a serious competitive response. A single-core Athlon II chop might be just the ticket; not quite as small a die size as the single-core (or if you prefer, dual virual-core courtesy of HyperThreading) Atom variant, but (like Via’s Nano) with features Atom can’t match, such as out-of-order instruction execution.

AMD also took advantage of this week’s Computex show to unveil the $102 Phenom II X2 550 Black Edition. As its name implies, this is a quad-core CPU with two cores disabled, translating to 7 MBytes of aggregate L2+L3 cache. It’s also built on 45 nm fabrication technology, therefore raising some doubt in my mind as to the yield robustness of the company’s latest-and-greatest process (note the italicized emphasis in the previous sentence). It runs at 3.1 Ghz, and its clock-unlocked status (along with its price tag) will be welcome news to extreme gamers.

Followup: Anand’s got benchmark results on both CPUs.