Saturday, September 22, 2007

On AMD Triple core

This comes from TeamAti's forum, and are my posts on AMD's Triple Core strategy. Basically one of TheInquirer's writers had written about AMD's Triple Core, and once again showed that nothing on the TheInq should be taken without a tub of salt. Part of TheInq's article concerned comparing coding techniques on the Xbox 360 as being similar to coding techniques on AMD triple core. My first response to the thread consisted of stating that I would have figured TripleCore would have more to do with Fusion than with anything else.


I don't see the connection between TC and Xbox 360...could someone explain this to me? - Shinigami

I don't think there is one. Specifically speaking, the Xbox 360 processor is based on PowerPC and uses technology derived from the Cell Multiprocessor architecture, namely the SPE's.

While AMD and IBM do share technology, such as SOI and APM, they generally work together on x86 and x86-64.

other thing to keep in mind is that the report came from The Inquirer, and the site has an earned reputation for being less than accurate. Both The Inq and the The Register are known for going out of their way to put excessive spin on a story... which figures as they were started by the same people. Sometimes I enjoy the over-the-top reporting and speculation, but it helps to keep a keg of salt on hand when reading.

I believe that the report is also factually wrong in developers already working with a Tri-Core system. The fact is, most x86, x86-64, and PowerPC is coded to make use of any threads available. Hence the reason why SMP enabled applications generally scale in performance as more threads are added to the available processing amount. Keep in mind that both Intel and AMD have been offering processor densities of 8 to 32 processors for several years, and both I think can be had in 64 to 128 processor densities.

Specifically this portion

This is a fact, since there are far more applications for Xbox 360 than native quad-core or more multithreaded apps

That is an outright fabrication on behalf of The Inquirer. There is a difference between software that is SMP aware, and software that is optimized for a specific number of available processing threads. The statement also ignores Linux on the Playstation3, which is a whole other matter entirely.

The problem is, much of the software that is built to run on the Xbox 360 is also being cross built to run on Playstation3, and x86 / x86-64. There really isn't a whole lot of optimizations available for triple-core PowerPC alone. There are, however, a lot of optimizations to make the applications SMP aware, and able to appropriately manage the available threads on each platform.


What does this have to do with 360? The 360 uses an IBM triple core. Are they going to start using AMD because the 360 has an ATi GPU? That would be a slap in the face to IBM then. - _[TeamATi]

Again... PowerPC... x86-64. They are
NOT the same architecture. I don't think AMD has any hands in PowerPC development. Microsoft could switch, yes, but the change from PowerPC to x86-64 would mean a completely new bus architecture, as well as a completely new AMD memory controller to handle the Yellowstone RDRAM in use.

Tri-core, is simply a Barcelona Quad-core, with on core disabled by a laser. This enables AMD to use chips that had only 3-cores operational, to be usable. - ColonelCain

Sorry. That doesn't make a whole lot of sense. Reason being AMD's A.P.M. (
Automated Precision Manufacturing) technology. AMD is fanatical about yield qualities. Sure, there are going to be some chips going through that have a forth die failure... but enough chips to make up an entire line-up?

Not out of AMD's fabs. Intel's fabs, oh yeah, I'd believe that story in a heartbeat.

As I understand it, with APM 3.0 a couple years ago whenever Fab 36 was coming online, AMD already had wafer level control during fabbing (
the process of making the chips) and was beginning to work on die-level control. Given that APM development hasn't stood still, there is reason to believe that AMD's current version of APM gives them complete control over the die during fabbing.

The fact is, low-end processors sell more than High-End processors. For every AthonXp that went out the door, there was probably 1.5 to 2 Duron's going out the door... same thing with Athlon64 and Sempron... and Intel Pentium and Celeron.

From a top down perspective, low end parts in processors, graphics cards, hard-drives, memory, power supplies... and everything else for that matter... make up the bulk of actual retail and OEM sales.

Given that Quad-Core is being positioned as a premium product, there are going to be quite a few more dual-cores being made than quad-cores, and sales history would indicate that Triple-Cores would sell more than Quad-Cores.

Ergo, AMD would either need a seriously high failure rate on their quad-core chips in order to fill the Triple-Core market... or the triple-cores are built from the ground up.

My opinion is that the triple core is an empirical test to insure that Fusion will work properly. Triple Core will give realistic information about power consumption, heat output, and voltage control. At the same time, it frees up room to start adding in a Fusion GPU.

AMD's history indicates that they plan ahead, and the idea that Triple Core is a setup for Fusion is a lot more in line with AMD's past than the line that quad-core yields are so bad as to generate another completely line-up.


Surprisingly after these posts were made on the forums, Kyle over at HardOCP came to the same conclusion I did about yield qualities... AFTER talking with AMD directly, which I hadn't.

HardOCP Editorial

However, Kyle did think up of something that I hadn't. Intel's Semi-Quad-Core processors are known to be poor overclocking choices, and Barcelona wasn't looking any better from initial reports that I saw.
“Would AMD rather sell a 2.5GHz Phenom Triple Core or a 1.8GHz Phenom Quad Core when hardly any piece of desktop software in the marketplace can actually utilize more than two cores? - Kyle

No comments: