[QuinEtiam] AMD FX/Bulldozer Registry Fix - 40% performance boost! - Page 23  

Originally Posted by Don Karnage

I have no idea why but this thread made me instantly think of this song

http://www.youtube.com/watch?v=umQCgaZnO0o

No i'm not high either...

Originally Posted by Dmac73

Funny that the pre release stuff for BD with extremely credible sources were deleted almost instantly, yet, this article is a no name website with NO source within the article what so ever, and is allowed to be posted in the news by a moderator.

Originally Posted by Xaero252

I see a lot of finger pointing and questioning...
Here is my current understanding:

In the past we have seen CPU drivers for windows (mostly XP, and mostly AMD.) This is NOT what is occurring here, its not some weird issue with the processor causing it (although it kind of is)

What is happening is this: The FX 8xxx series of processors has 4 cores, each core with TWO integer pipelines, effectively granting two processor cores for the price of one. (not "hyperthreading" effective, but almost literally additional core effective.)
This is all fine and dandy, except that when Windows (or Linux) goes to issue work to a core in the processor, it picks the core with the least load, and distributes things using a task scheduler. The problem here is that when a single pipeline on a bulldozer core is being utilized both cores are effectively "in use" even though Task Manager sees the cores appropriately, the task scheduler does not. This isn't really anybody's fault, but rather an oversight.
Standard CPU design from a by-the-book standpoint would never add a second integer pipeline to a single core. So nobody ever really thought to code a task scheduler that worked in that situation. So basically, AMD thought outside the box, and Windows got confused.

Oh not to mention that if the scheduler tries to run non-integer work on one of the integer only threads, the CPU returns an error and the command is re-queued and has to sit in line waiting to be executed again. Basically its a scheduler issue, and could definitely cause some HUGE impact in synthetic benchmarks and normal use. And even more issues in a cluttered server environment.