Replies to post #3264 on Advanced Micro Devices Inc (AMD)

[dougSF30]

wbmw, heard from whom?

Doug


http://www.theinquirer.net/?article=9140

...

This got tiring, so I decided to do something about it. At the Opteron launch, I tracked down an engineer who actually did some of the design of the chip, and asked him if the design was all new, or simply a warmed over design. The results, after a few clarifications, were undeniable. The changes were numerous, and sweeping, and not at all visible to someone with access only to the block diagrams and simplified PR handouts.

If you notice the reviews, all of them have the same AMD supplied architectural drawings, from block diagrams, to die photos, to pipeline stages. They give you a set amount of information, and it pretty much ends there, unless you are willing to dig deeply.

When I dug, here is what I found.

Functional units - "Ripped out" and redone completely
Decoders - Redone
Branch Predictors - Redone
Memory Management Systems - Redone
Overall, everything was pretty much ripped out and done over. The view from above may look the same, but there is precious little there that is.

To quote this engineer, "So while a high level block diagram of the K8 core looks like K7, I can assure you it is not "the same".

At that point, the PR side of the business stepped in, a gag was placed around this fellow's mouth, and someone put a striped square thing filled will yellow goo into my hand, as I was led off. µ

[fyodor]

wbmw: As far as I've heard, the Opteron uses the exact same backend as the Athlon. That includes the execution units.

It does have two more pipeline stages, but they are likely fairly early in the pipeline (maybe an extra decode / reorder stage or something), but I don't recall seeing any hard material on this.

Presumably, the execution units would also have to have wider data paths and buffers.

But, yes, all in all not much of a difference there.

Of course, the funny thing about all this is the time-line. A feature like prefetching was reportedly originally developed for Hammer, then adapted to the Athlon core when Hammer was delayed and Intel pushing hard. The improved branch prediction falls in that catagory too, I believe.

-fyo