TR Forums

Hi folks. I haven't read a Bulldozer article specifically saying how long its pipelines are. Even AMD hasn't officially publicly stated it. Anyway, the other day I was using my tablet to read some articles and I came across an interview with some AMDers. Sorry I can't post the link because I didn't bookmark it, but there, Mike Butler, Bulldozer's chief architect, said that BD's pipe is approximately 25% longer than K10. We know K10 has 12 main pipeline stages (by main, I mean, general integer instructions; FP takes longer to cruise through the pipes), and 1.25 x 12 = 15. I've had some slide images of AMD's other architecture, Bobcat, for quite a while now and what Mike Butler said got my mind racing. I Immediately looked at the Bobcat slides, specifically about its pipelines. Take a look:



From the very start I've always suspected that Bulldozer and Bobcat's pipelines aren't all that different judging from early slides showing their microarchitectures at a very high level. And if you look at the picture above (I believe it's from AMD themselves), you see that the main pipeline is 15 stages long. 15 stages isn't exactly a mile and a half long, but it's the longest pipe AMD has ever created. What do you guys think?

Edit - I think, looking at the slide, that all stages associated with Integer, FP and AGU, total 15 (I would think the arrow from RegRead to AGU is pointing the wrong way). The lowest portion shows what happens after an AGU operation requests something from the caches. DC1/DC2 probably mean Data Cache 1 and 2.

I'm missing something because that slide says 12.

Flip-mode, follow the arrows.

Yeah, some of the stages are consolidated. while i find this logic level interesting, lately i've been wondering why thread schedulers in modern OSs are so unkind to BD.
Is it difficult to change the order in which threads are parked on cores? could there be some on-die logic to help with this eventually?

i had also heard MS developed and then pulled a patch that improved BD's performance in Windows significantly.

Your memory's slightly off.

They developed a patch that helps Bulldozer support some, and due to some error it was released too soon. MS yanked it(I'm not sure why, maybe it wasn't fully tested), and then about a month or so later the patch was properly released.

An OS patch is not going to help Bulldozer, or any CPU arch by that much in overall performance.