PPD Predictor Thread (prelim)

[Vapor]

I know I've fallen behind on updating....promise I'll get it to done in the next few hours

Looks like the database is getting filled out well

[jcool]

Radaja, that is VERY high granted credit for a 920 at that speed. Which projects do you run?

[radaja]

the first few days i ran all of them and two days ago i changed it to HCC,HFCC and HCMD.
but today it finally started doing those three.i had set my thing to cache three days of work.

[Vapor]

9 PPD/MHz is pretty darn good for a Bloomfield at that speed (PPD/MHz does decline slowly at higher MHz). skinnee has a stock E5520 at 9.2 PPD/MHz whereas jcool's 920's (at 3800) are doing 8.3-8.4 PPD/MHz. Probably boils down to projects run and secondary settings (uncore and RAM namely).

A further investigation into the two settings you guys run might be pretty fruitful

[radaja]

i manually update throughout the day also.so maybe this has something to do with it?
after i have about 6-10 WU 's finished i send them away?

[Vapor]

I don't think that would have anything to do with it.

I think it's projects being run....I have my 930 at 3850MHz with a scant 2800MHz uncore and RAM at 1400 9-9-9-24 and I'm doing 8.8-8.9 PPD/MHz. I run only HCC.

EDIT: I'll switch my 750 to a new account and enable it for all non-HPF2 projects and run the PPD/MHz equation on the various projects after a few weeks, hopefully that will shed some light on which are the most gainful for PPD

[radaja]

yea it must be the projects because nothing special here with the overclock.
21x200@4200 HT on uncore at 3200 1600 6-9-6-21

[Angmaar]

This is an awesome chart. What would be cool to add is watts used vs. PPD vs. MHZ or something like that.

[Vapor]

This is an awesome chart. What would be cool to add is watts used vs. PPD vs. MHZ or something like that.

Not possible....there's a non-linear response between watts and MHz (considering a voltage increase accompanies most clock changes on this forum...assuming max MHz scales linearly with voltage, you actually have watts scaling with the cube of MHz when overclocking past stock voltage! [not a hard-and-steady rule, just a quick rule of thumb]) and there's also secondary components, individual differences between CPUs, and so so much else that just muddies it to impossibility.

If you want the most efficient cruncher possible...undervolt, overclock, run minimal secondary components, etc. When it's all said and done, it's very likely the most efficient is either a P55 i7 or a Gulftown system, but that's just my educated guess

[pirogue]

Here are a few more for you:

980x - 4.0GHz - Win 7 - x64 - daily usage
5/5/10 0:012:04:34:59 49,198 58
5/4/10 0:012:03:48:28 48,633 58
5/3/10 0:010:01:30:58 39,664 52
5/2/10 0:014:17:42:24 58,747 77
5/1/10 0:009:19:42:46 39,395 50

980X - 3.33GHz - Win 7 x64 - dedicated (haven't had time to tweak yet)
5/5/10 0:011:22:29:13 40,160 51
5/4/10 0:013:08:18:05 44,754 57
5/3/10 0:011:06:07:15 37,915 53
5/2/10 0:018:14:13:58 62,403 71
5/1/10 0:017:22:35:07 57,276 70

Phenom II - BE550 - unlocked 4 cores - 3.2GHz - media center light usage
5/5/10 0:003:14:42:18 11,475 16
5/4/10 0:004:07:04:43 13,756 19
5/3/10 0:003:18:49:01 12,143 14
5/2/10 0:004:22:41:33 15,426 19
5/1/10 0:004:16:43:36 14,963 18

[Vapor]

Massive chart, eh?

[jcool]

I don't think that would have anything to do with it.

I think it's projects being run....I have my 930 at 3850MHz with a scant 2800MHz uncore and RAM at 1400 9-9-9-24 and I'm doing 8.8-8.9 PPD/MHz. I run only HCC.

EDIT: I'll switch my 750 to a new account and enable it for all non-HPF2 projects and run the PPD/MHz equation on the various projects after a few weeks, hopefully that will shed some light on which are the most gainful for PPD

The 920 was running 200*19, 3200 Uncore, 1600 9-9-9 Triple Channel DDR3.
Running HCC and HFCC, like all my machines do.

[D_A]

Average scores over the last 18days:

Opteron 185
2.6GHz
5103 average PPD
Linux 64bit (daily drive)

Opteron 170
2GHz
3763.9 average PPD
Linux 64bit (boy's daily drive)

Dual Sossaman
2GHz
6301 average PPD
Linux 32bit (also running caching web proxy, content filter and DNS forwarder for whole network)

Vapour: ... how did you work out the 4.0 for jcool's Sossa? The numbers in his post give me 0.99PPD/MHz From your first example you appear to be multiplying by the number of threads. If I do that, my Sossaman scores a 12.6 on your chart.

[Petra]

Vapour: ... how did you work out the 4.0 for jcool's Sossa? The numbers in his post give me 0.99PPD/MHz From your first example you appear to be multiplying by the number of threads. If I do that, my Sossaman scores a 12.6 on your chart.

4.0 x 2000MHz = 8000ppd ...and that's about what mine has been putting out per calendar day for the past 4 or 5 days (it's new, numbers are in post #48). So, 4ppd/MHz doesn't seam unreasonable to me.

If 8000 points were produced in one calendar day, it would show up as ~4 CPU days on a sammy... so you'd be looking at 1.0 points-per-CPU-day/MHz, or 4.0 points-per-calendar-day/MHz, no? I'm way too tired to be doing this....

[meanmoe]

Great chart... I'm a little upset with the Thuban numbers though. I expected higher.

From this chart, I'm going to try and figure up 'best bang for buck' for my next machine. (Sossaman looks difficult to beat)

I don't know how many people have a killawatt meter, but it would be interesting to get power consumption for these as well, like Angmaar said. I guess the problem would be that it's not linear w.r.t vcore and vcore is not linear w.r.t MHz.

[Vapor]

Average scores over the last 18days:

Opteron 185
2.6GHz
5103 average PPD
Linux 64bit (daily drive)

Opteron 170
2GHz
3763.9 average PPD
Linux 64bit (boy's daily drive)

Dual Sossaman
2GHz
6301 average PPD
Linux 32bit (also running caching web proxy, content filter and DNS forwarder for whole network)

Vapour: ... how did you work out the 4.0 for jcool's Sossa? The numbers in his post give me 0.99PPD/MHz From your first example you appear to be multiplying by the number of threads. If I do that, my Sossaman scores a 12.6 on your chart.

jcool's averages ~8000 points per 96hr of runtime....which comes to 4.0 PPD/MHz. Yours comes in at 3.15 PPD/MHz, but you're running a bunch of other stuff on it, so its numbers will be diluted.

I'm not surprised by Thuban...it's actually doing better than 1.5x a Deneb X4, but that could just be because of the small sample sizes and the taint of non-dedication.

Of all the numbers I think are undershooting real performance, I think it's Clarkdale. I'd expect Clarkdale to pull 4.1-4.5 PPD/MHz on a dedicated system (based on what other Nehalem derivatives do for PPD/MHz/core). The 3.8 PPD/MHz on the chart is from a single system used by multiple users....I hope I can get more data on Clarkdale soon

In terms of cost effectiveness, a Sossaman is going to be hard to beat ($120 for CPUs+board+ram+HSFs?), but that's largely because it uses outdated parts. Microcenter's occasional deal of $99 for board + Athlon X4 would vie for that title as well if you can find cheap RAM. As for power....it's something that's way too difficult to chart for all of this, nevermind obtaining comparable data. It's a project I'd be interested in seeing the results of, but it would have to be done by a single individual that knows scientific method (and overclocking and computer and electrical engineering fundamentals) that has a lot of time and doesn't mind sacrificing production to accurately chart low-clock, high-efficiency output.

[Petra]

It's a project I'd be interested in seeing the results of, but it would have to be done by a single individual that knows scientific method (and overclocking and computer and electrical engineering fundamentals) that has a lot of time and doesn't mind sacrificing production to accurately chart low-clock, high-efficiency output.

I'm not volunteering for this...

[Vapor]

I have Lynnfield, Bloomfield, and Thuban here....but I'm not volunteering either. Too much of my watercooling/TIM testing to do using Bloomfield/Thuban....and my Lynnfield is already running a WCG test session to see which projects are the most gainful in terms of PPD/MHz.

[meanmoe]

I'll see if I can replace the Clarksdale and set it aside for a few days... This is not going to be an easy sale .

Whats the minimum sample period, in days, that you would need?

[Vapor]

3.5 is the lowside of what I'm comfortable with....5 is satisfactory. I don't think you should set it aside though....with the quorum points system, it'll be really hard to tell when 'dedicated' kicks in (unless you run it for like 10 days dedicated, which is too much to take away from the people that actually use it ).

I'm sure there's another Clarkdale around here somewhere

[D_A]

jcool's averages ~8000 points per 96hr of runtime....which comes to 4.0 PPD/MHz. Yours comes in at 3.15 PPD/MHz, but you're running a bunch of other stuff on it, so its numbers will be diluted

8000 points / 96hrs(4days)/2000MHz = 1 (based on calendar days, which my averages are)

I may have misunderstood how you wanted the original figures though. My numbers based on hours runtime instead of calender days are:

Attachment 104018

"new-bruce" is the Opty 185 @ 2600 linux 64bit
"boyz-desktop" is the Opty 170 @ 2160 linux 64bit
"sausageman" is the Sossa @ 2000 linux 32bit
"girlz-desktop" is a XP2600+ @ 2133 linux 32bit
"node1" is a XP 3000+ @ 2166 winXP home 32bit
"iCom" is a XP 2400 @ 2000 linux 32bit

[pfm3136]

I'm sure there's another Clarkdale around here somewhere

i3-540 @ 4.0GHz (HT on), Windows 7 x64, Boinc Manager 6.10.43 X64 (daily usage)
0:075:19:19:56 - 306,357

i7-860 @ 3.6GHz (HT on), Windows 7 x64, Boinc Manager 6.10.43 X64 (daily usage)
07-05-2010 0:006:06:25:02 25,449 51
06-05-2010 0:006:15:14:17 26,615 47
05-05-2010 0:007:21:52:04 30,787 53
04-05-2010 0:008:03:17:43 30,082 40
03-05-2010 0:010:18:00:23 40,258 47
02-05-2010 0:008:12:53:07 33,624 48
01-05-2010 0:005:22:01:05 24,238 40
30-04-2010 0:007:20:56:37 31,759 46
29-04-2010 0:010:02:27:39 38,003 71
28-04-2010 0:007:20:47:17 29,004 53
27-04-2010 0:006:17:47:00 24,799 42
26-04-2010 0:005:19:59:28 22,038 40
25-04-2010 0:008:19:11:09 30,457 48

Q6600 @ 3.2GHz, Windows 7 x64, Boinc Manager 6.10.43 X64 (daily usage)
0:072:20:08:47 - 273,957

[sRHunt3r]

Here are my main 3 crunchers, they are all running Win7 64 and boinc 64:

i7 920 @ 4.5 HT on


i7 920 @ 4.2 HT on


Q9550 @ 4.0

[soya_crack]

Are those dedicated or do you use them for other stuff, too?

[INFRNL]

Here are my main 3 crunchers, they are all running Win7 64 and boinc 64:

i7 920 @ 4.5 HT on


i7 920 @ 4.2 HT on


Q9550 @ 4.0

what projects are these running? I never get consistent results like that. I run HFCC & HCC. each project on a different machine. I get anywhere from the mid 20's to in the 40's I do not manually update though, just let the system run. these 2 machines are not used very much either; just some occasional surfing, but I do run gpugrid
edit: I have 1 920 and 1 w3520 ea 2 4ghz.