AMD's FX Piledriver performance in W7 vs W8.1 / W10

[TRWOV]

We've know for quite some time that the Bulldozer architecture that AMD introduced on their FX processors hasn't attained the performance that people hoped for, but what if our OS choice hurts the platform? could a simple OS switch make these CPUs shine?

These test were realized with this hardware configuration:
- FX6300 @3.5Ghz with CnQ and Turbo core disabled. Just in case, Power Options were set to High Performance
- Gigabyte GA-880GM-USB3 mATX board
- 2x2GB DDR3-1600 CL9 (RAM limited to 4GB as to not give x64 versions more legroom)
- HD7770 1GB to avoid tapping into system memory
- Silverstone 450w SFX 80+Gold PSU

Tested OSes:
- Windows 7 Pro SP1 x86
- Windows 7 Pro SP1 x86 with the FX Hotfixes applied
- Windows 7 Pro SP1 x64
- Windows 7 Pro SP1 x64 with the FX Hotfixes applied
- Windows 8.1 x64 (didn't have a copy of the x86 version)
- Windows 10 Preview x86 build 9841
- Windows 10 Preview x64 build 9841

For the drive space comparison the following setup was done:
- Hibernation disabled. This will delete the hibernation file (hibernation file size = available system RAM)
- Set a 2GB page file (file page sizes vary in Windows versions)
- Ran Windows update, installed all the critical updates and drivers
- Installed .NET 3.5 (for IBT)

Benchmarks used:
- Cinebench R15 (x64 only)
- Cinebench R11
- IBT with patched Linpack
All benchmarks were performed after a restart and waited for a full minute to make sure that all the background services and drivers had been loaded as to affect the benches as little as possible.


First comes Intel Burn Test. From left to right, top to bottom, W7x86, W7x86 with hotfix, W7x64, W7x64 with hotfix:

Whoa!!! Hold your horses!!!! A >2x increase in performance just by using an x64 OS? This seems to mirror some findings I got while crunching:

Another thing we learn is that the Hotfixes don't seem to do anything but we're using all threads there, maybe an uneven number of threads will show if the patches work?

From left to right, top to bottom, W7x86, W7x86 with hotfix, W7x64, W7x64 with hotfix:

Well, they do something, but not much. A more interesting development is that by using the shared FPU as a single unit we get >90% of its performance.

Now come W8.1x64, W10x86, W10x64:

We get a slight improvement by going with W8.1 but W10 actually seems worse, although not by much. The difference is within the margin of error so I'll let it slide but it seems weird that both the x86 and x64 versions showed this. Again we see a divide between the x86 and x64 versions. Clearly AMD optimized these chips for x64 enviroments.


Now comes Cinebench R11. From left to right, top to bottom, W7x86, W7x86 with hotfix, W7x64, W7x64 with hotfix:

Something weird happens here too: For some reason the fps on the OpenGL benchmark are better on x86 In this case the FX Hotfix actually hurt performance on the x64 version


OK, let's make a quick run with 3 threads. From left to right, top to bottom, W7x86, W7x86 with hotfix, W7x64, W7x64 with hotfix:

Again some erratic results, do the Hotfixes help or not? We see that the integer cores scale as they should giving roughly half the points for using half the threads.


Well, so far we've learned two things:

1- Using a 32bit OS is a deathspell for the FX CPUs.
2- The Hotfixes for W7 don't seem to offer a consistent improvement.

But what about using 32bit applications in the x64 version? Do these run gimped too?

From left to right, top to bottom, W7x64 running CB64bit, W7x64 with hotfix running CB64bit, W7x64 running CB32bit, W7x64 with hotfix running CB32bit:

Well, they seem to take a small hit but could be due to run variance.


Now comes W8.1x64, W10x86, W10x64:

Scores seem largely unchanged from the W7 results. Maybe a more up to date benchmark will show a difference?


Cinebench R15 to the rescue!!! From left to right, top to bottom, W7x64, W7x64 with hotfix, W8.1x64, W10x64:

(something I noticed while running CB15 on W10: CB identifies it as W8 for some reason )

Heck, not again. Windows 7 with the unpatched scheduler actually gets a better score. The W8.1 dip is strange too Maybe some heavylifting was going up in the background and I didn't notice?


So what did I learn after all of this OS shuffling?
There's no magic bullet that will unleash more performance from FX CPUs, specially on the FPU front. The integer cores do a good work and scale pretty well, just avoid using these on a 32bit OS.



Corollary: SSD early adopters rejoice!!! That old 30GB SSD is relevant again!!!!

 

[Steevo]

Nice writeup.

And yes, at the end of the day AMD's design flaw isn't the shared resources, its the hugely slow caches, it hinders performance in X87 among other things.