Let's talk Opteron 6100 (K-10) vs Interlagos! 6300

[GPU timpster]

Hello,

I'm wondering if using a "true" (not six modules) 12 core Opteron 6100 series would outperform an Interlagos (6300) series with eight full "modules" split into 2 cores for each module. I believe that with the Bulldozer arch, if I were to only use half of the module (1 core from a module) I could not access the full power of it. I'd have to load the chip evenly to use it's full power.

With the older K-10 like Phenom II x4 and x6 chips, you don't have "modules" but full cores without the extra threads.

Would having 24 full cores outperform 16 "modules" or is the addition of the new CPU features like 6200 and up having AVX as well as the 6300 series having better FPU performance going to take over any advantage of extra cores?

Here's some history on K-10:
http://www.hardwaresecrets.com/inside-a ... hitecture/

Here's some history on BD: https://en.wikipedia.org/wiki/Bulldozer ... dozer_core

On the WIkipedia page, there are tests that show the K-10 sometimes outperforming the bulldozer architecture, but I'm wondering if this is applicable in F@H. I can get 6100 opterons for quite cheap now, so if they will perform similarly, I'd be surprised, but I don't think they would.

[Nathan_P]

These days its a waste of cash unless you already have the machine left over from the Bigadv days. If you really want to try it out then I would say the 6300's would be best as they have the best ipc and avx. There are lots of threads on the AMD 6xxx chips on the various team forums, 5 minutes with google will give you a big list

A dual E5 v2 setup will get the same points for much less power draw.

[bruce]

In the context of this forum, your word "outperforms" is probably a specific reference to FAH's performance, not a more general measure of performance.

In fact, FAH for CPUs uses a preponderance of floating point arithmetic and generally speaking a module shares the floating point execution hardware between two threads (on what was formerly called a FPU), so there's some loss of FAH's performance. Also, as a general rule, most non-FAH programs don't use any floating point arithmetic, but rather rely on fixed point arithmetic (or what was once called an ALU), and each module has enough ALU hardwre to process both threads without degradation. Thus performance of non-FAH programs is not reduced by the sharing -- so there's a good reason for hardware designers to create modules (Certainly they're cheaper.) -- though for FAH (only), having full cores would be an improvement.

Let's suppose that a FAHCore contains 80% FPU operations and 20% ALU operations. If one FAH thread happens to be sharing a module with a non-FAH thread, the FPU would be 20% idle (neglecting many other factors) during the 20% of the time that it's being limited by the ALU.

Now, suppose that two FAH threads happen to be sharing the SAME module. Each would use 50% of the shared FPU and each would still use a significant percentage of the ALU. Nevertheless 50% + 50% is greater than 80%, so FAH gets more done than if only one thread is using the module but not as much as if those two threads had access to dedicated FPUs.

Remember, I'm talking in generalities so i don't pretend I'm being 100% accurate, but you do get the idea.

[GPU timpster]

Wow, this has been very helpful. Thank you both Bruce and Nathan P. I'll have a look around on other forums too, but so far I think the "new" FPU performance in the 6300 would be more beneficial for FAH. I appreciate the heads up on the Intel Xeon performance, but I don't feel like a can make a 4P system for a reasonable price. The biggest factor is a case, and I don't want a rack mount case.

[Nathan_P]

For the xeons I was only talking about a 2p system, 4p xeon is stupid money. 2p boards for either AMD or Xeon can be had in EATX size with then means any reasonably large tower case