SMP 10 > 12 !?!

Well, today for a while I ran out of GPU WU's for my 5850 for some reason so I thought I would have another play with at running SMP 12.

I roughly tested this a little some time ago and did not loose any PPD changing from 12 down to 10 cores for SMP folding.

Today's SMP WU - 7083 - was a little boring with approx a 13 hour WU giving only somewhere around 16k PPD while running SMP 10 alongside my GPU 5850 folding. I figured I might get up to 19 or 20k changing up to SMP 12.

But... :

Xeon X5650 6 cores / 12 Threads OC'd @ 3.8GHz

Running SMP12 (settled down to the following improved TPF after 6 frames worth while not using the computer for anything):
TPF = 9:00
PPD = 13,109.8

Running SMP10 (Best):
TPF = 7:30
PPD = 17,238.5

Running SMP10 (Avg):
TPF = 7:38
PPD = 16,256.6

(testing done with no GPU folding --- SMP 10 result is exactly the same With GPU folding running - I have 7% CPU time free)

Interesting that reducing the folding cores by 2 Increases PPD quite substantially
and also nice that it leaves extra CPU time for daily use of PC (even when feeding my Radeon 5850 @ 5870 speeds) without interfering with my PPD ^^

Maybe people running 8 core chips will get better PPD from SMP 6 also ?
I have seen people say they get better results from SMP 8 than 6 before though.
Worth checking sometimes though it seems

Your ppd is real close to what I get with 8 core @ 2.66Ghz.

It is highly likely that trying to squeeze heavy threads onto "not busy" HT cores not only slows down the threads running on HT, but also gets in the way of the non-HT threads on the same core.

I think this has to do with the fact that HT works by allowing extra threads to run when the primary thread has idle time. Since F@H is pretty efficient and likely uses almost all of the cores resources the HT thread will keep getting bumped. Adding 2 more threads is likely reducing performance because now the scheduler is trying to smash even more through the CPU causing a bottleneck.

Yeah indeed. I knew HT'd cores rely on the Physical cores to do their floating point calculations.
I have seen from before the no performance drop going from SMP 12 to 10
But was not expecting such an increase in performance for doing so

This 7083 is a horrible WU, maybe the lowest PPD one I have had in a really long time.
The other day I had a nice series (10083 to 10085) that game me about 36k PPD for SMP10 ^^
but average WU's get me about 24 - 26k I guess

That's quite suspicious - I think there might be another reason for this difference, which is really substantial.

Six cores, FAH says, will crash some wu's, entirely, so I'd recommend avoiding that, as well as any odd number of cores, for folding.

Where does it say that about six cores? After bigadv for eight cores was gotten rid of, I eventually reduced FAH on my server (i7 930) to only 6 cores, and the PPD seems to be much more stable than it was at 8 cores.

Adak said:

That's quite suspicious - I think there might be another reason for this difference, which is really substantial.

Six cores, FAH says, will crash some wu's, entirely, so I'd recommend avoiding that, as well as any odd number of cores, for folding.

Click to expand...

I was running SMP 6 on my Phenom II x6 for two years until a week ago when I replaced it. Never had any issues that were not obvious OC stability related. For a short time I was running SMP 5 to keep a core free also no problems. Where have you seen them say there are problems?

Maybe people running 8 core chips will get better PPD from SMP 6 also ?
I have seen people say they get better results from SMP 8 than 6 before though.
Worth checking sometimes though it seems

Click to expand...

At 3.8GHz on the x6 I was pulling 20,000 PDD. Looks like FAH is finally showing off the limits of Intel's Hyperthreading. So this is a 6 physical core chip with 12 threads possible? What do you get with just SMP 6? What about disabling HT in the BIOS and running SMP 6? Really just personal curiosity regarding the architecture. I'm not sure if HT splits the full pipeline of each core or just shares it, but perhaps FAH get better results from full use of a single core's pipeline rather than splitting the work over virtual cores and then competing for the pipeline.

On a similar note I'm wondering if I will get the same performance on my new AMD FX 8320 with SMP 4 as I do on SMP 8. The chip only has 4 FPU's for the 8 integer cores although this shouldn't pose much issue for FAH as the FPU's can handle four 32-bit SSE instructions at the same time, so each "core" isn't competing for resources the same as HT does on an intel processor. Although the FPU could still cause a bottleneck in the right workloads.

At 4500mhz I see about 61-63k ppd from the CPU alone, depending on the wu. I don't think they are at their limits with ht, but if you interrupt f@h you are deducted the CPU time lost. It wants all of you're CPU to it self. I saw a significant decrease disabling ht.

have you checked your processor speed to make sure turbo is not kicking up your speed when your utilizing fewer cores?

I have disabled most anything that can cause a CPU speed change and while testing I was watching the clock speed carefully - sat steady at about 3799.8 MHz
Also while doing the above testing I had nothing more than Core Temp and Task Manager n basic Windows Start up stuff running

I have a few busy days on here now - fixing my R1 and taking on a new work project ^^
When I have a spare moment and a long WU to play with I might try some SMP 8 and SMP 6 with HT disabled

But looking at Task Manager the workload with SMP10 is more or less evenly spread across the cores so PPD should just drop from here on down...

Still there is nothing like some real world testing to be sure