FX8350 GFlops Drop?

OllieTheWolly · 13 Apr 2013 at 15:40

Hi guys, got a wierd problem here.

I've just replaced my AMD Phenom x6 1090T @ 3.6GHz/core with a new FX8350.

My system is stable, but according to Intel Burn Test and LinX, it performs at 35GFlops. On my Phenom I was running at 74GFlops.

I've stopped all the turbo technology and power saving features in the bios, so the core multiplier cannot change. Additionally I've installed the windows hotfixes KB2645594 and KB2646060, and am using all the up to date versions of IBT and LinX.

You guys got any ideas?

Specs:
Asus Crosshair V Formula (Bios 1703 - supports FX8350)
16Gb 1600MHz Crucial Ballistix Ram @ 8-8-8-24
VTX3D Radeon 6950
750W PSU.

NB @ 2200MHz
HT @ 2600MHz

mmj_uk · 13 Apr 2013 at 20:44

Check for any signs of thermal throttling. Or it could just be that IBT is not optimized for AMD FX.

jcypser · 13 Apr 2013 at 20:54

I just ran the Intel Burn test and got 35/34 Gflops myself.

My Spec
AMD FX-8350
Asus Sabertooth 990thFX R2.0
8GB Corsair Vengeance 9-10-9-27
ATI HIS HD 7950 IceQ BOOST 3072MB
750W psu

mmj_uk · 13 Apr 2013 at 21:17

I don't think it's optimized then, iirc that's why I ignored it as a stability test because the temp was much lower than Prime95 (which is a sign that the CPU is not being kept busy).

To judge performance you are probably best off using Cinebench, Prime95 for stability/stress testing.

Something Else · 13 Apr 2013 at 23:36

I get 125 Gflops on ITB but thats with a 2500k at 4.7

SlugForAButt · 14 Apr 2013 at 00:34

Somthing Else said:
I get 125 Gflops on ITB but thats with a 2500k at 4.7

Well that was worth posting...

Futures · 14 Apr 2013 at 07:30

HT @ 2600. Try putting it back to stock speeds I think its 2100 for piledriver? I know with the phenomII increasing HT had a negative impact, maybe the same for piledriver.

Hagar · 15 Apr 2013 at 09:34

joeyjojo · 15 Apr 2013 at 09:58

^ Something weird going on there, the results column is screwed up.

Hagar · 15 Apr 2013 at 11:05

joeyjojo said:
^ Something weird going on there, the results column is screwed up.

This is documented, particularly with high clocks on AMD processors running IBT. It may be a cpu register which returns a NAN (not a real number) ie 1/0 or it may be a program function similarly copping out.

The program was running perfectly as in the IBT runs completed, all results in a tight range and the CPU was not throttling. Stability was tested in prime for 20 minutes before running IBT so temps were not excessive although toasty. HWmonitor was used to monitor the CPU temps and CPUz the CPU frequency.

This run was only to get a 100Gflops result and not to prove anything else. It did provide a lot of info on the fact that all out CPU speed was not the answer to gflops, but a combination of memory, CPU and HT/NB was the way to go.

OllieTheWolly · 15 Apr 2013 at 11:52

Thanks for the posts guys, and sorry for not being back in 2 days

Futures, I've tried with everything at stock (HT @ 2200MHz) and overclocked, and my GFlops vary between 35 at stock to 42 at 8 * 4.5GHz.

I've found a forum which discusses a similar issue (http://forums.anandtech.com/archive/index.php/t-2289809.html) and they seem to suggest that using an FMA optimised version of Linpack (as opposed to AVX) is the answer. Does anyone know where such a thing could be found - google returns lots of info but no executables or anything...

This could make sense given nkata's results above - I note that FMA3 & FMA4 are not part of your instruction set. I am currently replicating (as best I can) you system from what you've posted, and this is what I've got:

NB @ 2400
HT @ 2400
FSB @ 300
CPU @ 15*300 = 4.5GHz
16GB RAM @ 1599MHz, 9-9-9-24

This combination gives me a max GFlop value of 42.8.

I've performed a few other benchmarks which seem to suggest that the CPU offers increased performance over the Phenom x6. Cinebench R11.5 rated the CPU with a respectable 7.39 (all cores) and my own heavily threaded semiconductor simulations are around 30% faster.

I've also found a few other people who get around 35 - 40 GFLops on the FX8350, so my guess is that it's a problem caused by Linpack.

Hagar · 15 Apr 2013 at 12:35

http://www.datafilehost.com/download-b6f1cf10.html

Use this patch.

uncheck use our download manager and click download.

EDIT; my memory was at 2GHz, I think that makes a difference in IBT tho' probably not much in the real world outside benching.

wolvers · 15 Apr 2013 at 12:49

You're using the wrong version of IBT, that's all. The one nkata has linked is the correct one for Piledriver.

OllieTheWolly · 15 Apr 2013 at 13:45

Thanks guys! Using that patch I now get 93GFlops (thats with all cores at 4.5GHz), representing a 25% performance increase from my Phenom.

mmj_uk · 15 Apr 2013 at 17:04

Use Prime95, it's still a better stress test.

http://forums.anandtech.com/showpost.php?p=34536008&postcount=593

Phixsator · 15 Apr 2013 at 17:27

OllieTheWolly said:
Thanks guys! Using that patch I now get 93GFlops (thats with all cores at 4.5GHz), representing a 25% performance increase from my Phenom.

I just cant get my results near your 93 gflops @ 4,5ghz. Installed the AMD Optimized LIN package and with that at 225x20 i get around 74 gflops.

Something Else · 15 Apr 2013 at 17:31

SlugForAButt said:
Well that was worth posting...

Irony.

But I aim to help.

OllieTheWolly · 15 Apr 2013 at 18:35

Phixsator said:
I just cant get my results near your 93 gflops @ 4,5ghz. Installed the AMD Optimized LIN package and with that at 225x20 i get around 74 gflops.

It could be an FSB issue - my 93GFlops was at 15 * 300MHz. But that said at 22.5 * 200MHz I got 89.9 GFlops.

How much RAM are you using for LinPack? I get 93.1 GFlops when I use 14GB of RAM, but I get 74.6 GFlops when using 256MB.

Also, something I noticed when I first installed the CPU, was that when running linpack the CPU multiplier would keep changing as part of the energy saving technology, and spent about 20% its time around 17.0 (stock was 20.0) which severely reduced my GFLops rating compared to when I had all all that disabled. I think it was C6 and C1E or something that had to be disabled...

I guess you will have already disabled them but I have no other idea why your rating would be lower otherwise. (Unless you've disabled Load Line Calibration or can't supply enough power)

Phixsator · 16 Apr 2013 at 13:59

OllieTheWolly said:
It could be an FSB issue - my 93GFlops was at 15 * 300MHz. But that said at 22.5 * 200MHz I got 89.9 GFlops.

How much RAM are you using for LinPack? I get 93.1 GFlops when I use 14GB of RAM, but I get 74.6 GFlops when using 256MB.

Also, something I noticed when I first installed the CPU, was that when running linpack the CPU multiplier would keep changing as part of the energy saving technology, and spent about 20% its time around 17.0 (stock was 20.0) which severely reduced my GFLops rating compared to when I had all all that disabled. I think it was C6 and C1E or something that had to be disabled...

I guess you will have already disabled them but I have no other idea why your rating would be lower otherwise. (Unless you've disabled Load Line Calibration or can't supply enough power)

As you pointed out it was the amount of ram i used in IBT. The second i turned it above the standard 1024 to something like 4gb+ results were much better giving me roughly 87gflops at 4,4ghz.