Kaos Kid wrote:
Tator Tot wrote:FX-8120 is $220 right now.
If you disable Core's 1-3-5-7 in the BIOS (leaving 0-2-4-6 enabled) you can get up to a 20% increase in IPC, Realistically, it's more like 15% increase in IPC.
AMD's biggest flaw with these chips is marketing them as 8 Core processors, when they're realistically 4 Core parts, with a hardware level hyper threading.
Is this for all applications, or only those times when hyperthreading wouldn't be helpful? Give some basic examples please, I had to wiki "IPC" just to see what the heck you were talking about
It's more of just in a general sense. When you disable the 2nd "core" on a Bulldozer module, you free up a decode & prefetch unit so it can operate for the other core. Like wise, you also shorten the pipeline; meaning that instructions take less time to be executed, speeding the whole process up.
Now, if you leave Bulldozer on "full blast" so to speak, with all 8 "Cores" enabled, then you can see instances where the Bulldozer architecture will shine. Specifically, when you have applications that take advantage of FMA4 & AVX Instruction sets.
Unfortunately, most programs barely go past the SSE2 instruction set (which is from 2001, almost 10 years old now.)
If you compare the performance of SSE4.2 on Intel & AMD based chips, you'll also see Bulldozer can take a lead; but not as great as before.
Another instance where Bulldozer can push ahead is with XOP & CVT16; which are two instruction sets derived from SSE5 (as SSE5 isn't finished yet.) Though, these instruction sets are really only going to be used by HPC systems, until SSE5 is finished.
Bulldozer's biggest downfall is AMD's marketing (like I said, Bulldozer is a 4 Core part. 1 Module = 1 Core. No matter how you look at it, from an engineering or physical stand point. Of course, AMD will argue this with me till no end, since they want to throw more marketing at the wind with their Opteron parts saying they've got the first 16 Core chip and what not.
The other downfall to bulldozer is that modern software is just not optimized for modern CPU design. The performance we see from both AMD & Intel right now, is increasingly getting "gimped" by Software that wishes to support old CPU's like the Pentium 4 & Athlon 64 generations.
Sadly though, FAR too many businesses are still using Windows XP and old Pentium 4 / Athlon 64 based systems, which hold back software development. If you were a software engineer you would easily understand what I'm saying; as you have to easily add in an extra 400 lines of code to simple programs like Web-browsers so that they actually work on those old systems. Back-doors, alternate paths, and in some cases, completely new code; just to support what was outdated 6 years ago? Seems insane right?
Now, don't get me wrong; Bulldozer does have some fundimental flaws when you actually look at the design of the chip.
Phenom II chips have 64KiB of L1 Cache per Core, a Bulldozer Module has 32KiB of L1 Cache split between two x86 cores (effectively making it 16KiB of L1 Cache per x86 Core on a Bulldozer module.) This right here, will kill IPC on simple tasks that simply don't hit the L2 Cache, let alone the L3 Cache.
LIke I said before, AMD's design of the Prefetch & Decode nodes inside the chip were also too far apart and too few, increasing the pipeline and slowing down the entire chip (since that's one of BD's shared resourced on a module.)
You can take what I say here with a grain of salt, since none of this is confirmed; but the re-iteration of Bulldozer will be the Piledriver Core, which has a redesigned Prefetch & Decode process to increase IPC on a Module by 10-15%. This will easily catch AMD up to Intel's Sandy Bridge design; but Intel will have Ivy Bridge out by then and Global Foundries 32nm Process has given them nothing but issues with Bulldozer.
What remains to be seen is if we'll have a small jump from Piledriver (E.G. GTX 480 to GTX 580 ) or a huge jump ( Phenom I to Phenom II ) from the Piledriver revision.
What I can see, is that any negative press you see about Bulldozer right now, is as granted as Fermi (GTX 480) got at launch; but it's also over hyped to an extreme margin.
On the other hand, AMD didn't release bulldozer completely deaf, blind, & stupid. They know what they were doing, and at least have a vague idea of where they're going to move in the future. If this pans out for them? Well dear god I hope so, as AMD's release cycle is gonna be fast-tracked with Bulldozer and it could give Intel a run for their money down to the 11nm manufacturing process (assuming that Global Foundries can keep up with AMD's demands.)