Modules vs Cores with data (AMD FX series)

[bob4933]

For the purposes of this conversation, I'd like to hold piledriver and bulldozer equivalent if thats ok.

Notes ->

All tests were complete 5 times with the exception of R15 single core, which was ran 3 times. Averages of scores were taken. CPU is an fx 8320 on a sabertooth 990fx, running 1600mhz ram at stock speeds (only multiplier was touched. Turbo, APM, and other power saving devices were disabled). Super pi ran WITHOUT any modifications.

I ran valley, but the scores wound up being erroneous as the data sets seem to be a much wider range unfortunately.

This is observational data only. Ive only seen singular data on core count,without consideration to the modules or architecture used on these cpu's.

8 cores 4 modules
---------------------
- Geekbench 3 32bit 2134 single core/11404 multi core
- super pi 1m 22.651 seconds
- R15 multi score 638
- R15 single core 95

4 cores 2 modules
---------------------
- Geekbench 3 32bit 2114 single core/6217 multi core
- super pi 1m 22.500 seconds
- R15 multi core 318
- R15 single core 95

4 cores 4 modules
---------------------
- Geekbench 3 32bit 2067 single core/6081 multi core
- super pi 1m 23.17 seconds
- R15 multi core 303
- R15 single core 90


Obviously the 8c/4m was simply a baseline.

Gains from 4c/4m to 4c/2m was ->

2.27% for single core geekbench
2.23% for multi core geekbench
2.97% for super pi 1m
4.95% for R15 multi core
5.55% for R15 single core


Again, purely observational data. Results weren't surprising, but most likely impossible to differentiate in day to day tasking. R15 CLEARLY favored the paired modules vs the independent cores. Im also under the impression, you are able to clock higher having 1 core per module, which would obviously offset the difference for sure.

 

[ShrimpBrime]

Hey this is awesome thanks!

Couple of Q's -

What exactly was the Cpu clock speed during these tests? Can you also give us NB and HT speeds.

Can you run these tests overclocking single components, such as NB stock vs OC, HT stock vs OC and do them one at a time?

Also, I'd be curious about a stock all system vs a stock system set up manually. Meaning just turning off green vs leaving it on and see what kind of effects it has.

In addition to my Q's for you, I was also wondering about RAM. I notice you used 800mhz. Most boards default to 667mhz. Could you compare these two settings and then Try FSB or reference clock (200+) to obtain 800mhz from stock 667 divider. This will make some interesting outcome. Of course back the HT and NB CPU multis to compare near stock speeds. Since this is for eye balling and reference observational data, I'd be interested in this tid-bit in particular.

The more data you add, the better the collective data starts to show it's strengths and weak points.

Again thanks. More threads like this help OCer's understand where performance is at.

 

[bob4933]

4000 +-1ghz

My board defaults to 800mhz dual channel, not 667. Not sure what you mean by that, sorry. Actually, it defaulted to XMP, is it supposed to do that? Or you're asking me to clock down the ram and try again?

NB and HT were 2200/2600 stock speeds


And I will try to accommodate your requests

 

[Johan45]

Is that a Sabo REV1 or something? I'm just curious how you are separating cores from modules.

 

[bob4933]

Registry changes in windows, can set active cores for programs. For others using the boards that allow individual core on/off.

 

[Johan45]

I wonder if there's a difference between one set in BIOS VS set in Windows. If I find some time I might be able to do a few tests but don't know. Things are really busy here ATM. It's been nearly 2 weeks since I have even had a board loaded on the chopping block. I do have the CHVF that can separate cores in BIOS

 

[EarthDog]

I wonder if these tests were really chopping off what was believed to be chopped off... Which one is a 'full' core and which ones are integer/floating point only (I can never remember which of these is not in a core, LOL)?? Which cores did you disable for this testing?

I would be interested to see results on the other side as well... If you disabled 1,3,5,7, try now disabling 2,4,6,8 and see what the results yield.

I also forgot to ask if the OPs windows is updated with the AMD Scheduler hotfix (always tries to use one core /module) already...


More of the same testing: http://www.xtremesystems.org/forums/showthread.php?275873-AMD-FX-quot-Bulldozer-quot-Review-(4)-

 

[Johan45]

You have to picture the module like one inlet with a divider( fetch,decode,fpu) and the cores. It's really hard for me to say whether the other core is inactive or just unresponsive. But I do know it was easier to run the core speed up when the front load resources weren't being shared so it has to be doing something even if both halves of the module are running the same speed.
One thing I do know is with my bios I can't be selective as to which core of the module I want active. It's only 1,3,5 and 7. If I select core one and 6 for example the whole module (5+6) is disabled so only core 1 shows in windows.

 

[EarthDog]

I was just reading up a bit more on the architecture so I can wrap my head around it better... http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2

Im confused... you mention you can only disable 1,3,5,7 but then say 6? What did I miss there?



I just wonder then if his software method, setting affinity maybe(?) could do that?

 

[Woomack]

If I remember it right then you could disable only 1/3/5/7 or whole modules. In this case we don't really know if core numbers in BIOS are exactly the same as we see in Windows and if setting manual affinity actually disables right cores. I'm not really sure about that.

 

[EarthDog]

If I remember it right then you could disable only 1/3/5/7 or whole modules. In this case we don't really know if core numbers in BIOS are exactly the same as we see in Windows and if setting manual affinity actually disables right cores. I'm not really sure about that.

That is exactly my point and why I asked...

I wonder if these tests were really chopping off what was believed to be chopped off... Which one is a 'full' core and which ones are integer/floating point only (I can never remember which of these is not in a core, LOL)?? Which cores did you disable for this testing?

I would be interested to see results on the other side as well... If you disabled 1,3,5,7, try now disabling 2,4,6,8 and see what the results yield.

 

[Woomack]

It's too confusing. No problems like that with Kabini

 

[Johan45]

Core "0" in windows is core 1 in BIOS. So we have 1and2, 3and 4, 5 and 6, lastly 7 and 8. What I was saying through bios is that only 1,3,5 and 7 can be active on their own. So for module 7/8 if I disable core 7 and enable core eight both are disable in windows. It's as if the bios treats the "first" core of the module as a switch and it must be active or the whole module is disabled. Follow me?

 

[EarthDog]

I do now. Is it just me or does it appear that some of the other testing was able to disable specific cores within the module?

 

[Johan45]

Not really if you look at his bios pic this shows what I was talking about and pretty sure this was how he tested since I'm probably using the same bios. It isn't widely available to my knowledge.

 

[bob4933]

If I remember it right then you could disable only 1/3/5/7 or whole modules. In this case we don't really know if core numbers in BIOS are exactly the same as we see in Windows and if setting manual affinity actually disables right cores. I'm not really sure about that.

Thats a great point. I'll definitely mess around with this as well. Busy this weekend, I'll try to get around to it sunday.