• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

Maximum PPD per Box

S

Sunin

[H]ard|DCer of the Month - August 2008
Joined
Dec 27, 2005
Messages
3,421
So I started working various configs today and thought it would be good we all share what has worked and not worked.

So far what I have noticed is with Affinity Changer active I get the most points by letting AC do its thing for the Dual SMPs and by setting Priority to High on F@H11 (GPU Client) and Affinity to all 4 cores. This has netted me after several hours some 1.5k PPD increase!

Went from 5.8k PPD to 7.3K PPD and sitll holding after 2 hrs of testing!

Well hmmm letting affinity changer do its thing ends up reassigning F@H11 to core 3 only after a while... so I disabled it and will let it run.

F@H11 (GPU)
-High Priority
-Affinity all 4 cores

F@Ha1 (SMP)
-Low Priority
-Affinity assigning 2 process per core



 
After now 4 hrs of running

Disabling F@H Affinity Changer is the way to go with 2 SMP's

1st Quad with GT = 7300 PPD
2nd Quad with GT = 7000 PPD

Each one added about 3900 PPD by adding the GPU to the mix.

280W / 7300 PPD = approx 26 PpDpW

Not too shabby! Its just a matter of getting them setup (So now I have 3 boxes all in the 7k PPD range! hehehe) Less boxes + more output per box = Sweet!

 
It's also my plan to buy 3 more 8800GT for my 3 other boxens (even my WHS box, hope the OS is not finicky with the driver).

 
7kPPD per box should be very sustainable. Especially if those ar 8800GT numbers you posted (it looks like they are).

I've got one main rig I fold with that's running a Quad@ 3.1 and a 9600GT that I picked up 4 months ago when 8800GT prices were still ridiculous. I started with the GPU client, one SMP client and a console client. Total PPD for this setup topped out at 5960 with the GPU providing a high of 3736 of those points...

I then switched to running two VM SMP clients and the GPU client and since this is my main rig, total PPD varies with my usage, but runs from a low of 6050PPD to a high of 6415.. :)

So 7k from an 88GT should be no sweat. Now wait till we can hookup 2 88GT's on a Quad to fully soak two cores, and a VM SMP should net between 10-11k PPD.. :eek: Time for me to start shopping I guess...

I'm wondering what effect running a GPU on a less-than 16x PCI-E slot will be. My XBX2 has 3 16x capable slots, but using multiple slots drops two of them to 8x and the third to 4x. Do you know of anyone running a GPU on less than a 16x slot? I'd think it'd slow down the client somewhat but not sure how much. I doubt it'd be a 50% decrease to match the link speed. hmmmmmm...

Anyway, nice testing!

0.png
 
Which OS are you running on and what drivers ??
I've read that Vista only needs ~10% of a core to run the GPU2 client on but XP needs 100%, thats running on the 177.35 drivers.
Thats one thing I'm going to test once I've got a new 9800GTX card.
Swop my box from XP to Vista and see if I can run the GPU2 client and 2x VM's on a Quad with only a ~5% slow down on the VM's.
Otherwise I'd recon on a ~15% slow down on the VM's under XP.

I need to set Affinity on my main Dual-Quad otherwise it lags like fun.
I use SMP Seesaw Pro to isolate one VM on 7 CPU cores.
Then I move the other three VM's and all the VM software that shows up in SMP Seesaw Pro onto the 7 cores .
This means the the 4x VM are the only thing on 7 CPU cores.
The GPU client and Windows run on the 8th CPU core.
The box is as smooth as anything for most stuff, just the smallest stutter when playing DVD's is the most noticable slow down.
Without any affinities set or it set the other way (VM's & Windows on 7 cores, the GPU2 client isolated ) it lags so much as to become unuseable, could not even open a DVD as the screen lag was to bad.

On my Dual-Quad it makes no difference PpD wise if I have affinities set or not, It just makes a difference in how useable it is.

Luck ......... :D
 
I'm getting ~8300PPD (per fahmon) using 2xSMP (set to 'idle', no affinity changer) plus one GPU2 (set to 'slighty higher'). Vista 64(hacked cuda inf) Q6600(3GHz) 8800GT(stock) 8GB

This is a "first run" so my average will likely change a bit with different WU.

edit for tigerbiten: CPU usage for GPU2 ranges from 8% to 25% of one CPU(1/16 of total CPU available) on the quad when running alone without SMP clients.
 
Tigerbiten said:
Swop my box from XP to Vista and see if I can run the GPU2 client and 2x VM's on a Quad with only a ~5% slow down on the VM's.
Otherwise I'd recon on a ~15% slow down on the VM's under XP.

Click to expand...

On my system (XP32 and 177.35 drivers using VMWare Server 1.05) it's the GPU client that takes the big hit... 35% (based on a 3684PPD rate for the GPU with a single VM SMP and one Console client vs. a 2388PPD rate for the GPU with 2x VM SMP clients). The SMP clients lose MAX 5% (VM SMP single client @ 2081PPD with GPU client vs 2x VM SMPs @ 2011-1977PPD)

GPU takes a bigger hit but overall boxen production is greater...

0.png
 
I'm runnin winxp 32bit, 177.35 drivers on all 3 machines.

Hope that helps... The GPU client is eating about 16-20% of one core

We really need a tool to be able to manually assign threads to cores and then it checks ever x minutes to verify... I think that would be the best if we could do something like that... I'm looking into maybe creating a batch file that I could setup an event that would take the PID and assign it to a specific core via command line every 5 minutes.

Just a thought...

 
Tiger, I have similar results a relic in 32 bit Vista with the newest normal drivers


 
I also get a similar PPD output with my main PC (Vista x64, 8800GTS 640MB and Q9450 at 3.5 GHz).

 
I'm guessing the fact that you can address all the memory and are running native 64 bit is helping this!... I might have to try with my water cooled setup to reinstall xp64 again... cuda just released beta 177.35 drivers for all operatings systems... well windows operating systems that is.

What are you all using to adjust affinities? or are you letting all processes run on all cores?

I just d/l'd seesaw pro.. very cool tool, but I'm curious is when the service comes back up after a turn-in will it reset the affinities?

Is there any tool that can perm lock a process to a set core?

Also I am still noticing that my 3.2Ghz machines are not doing as well with the SMPs

 
WinXP32 Q6600@3.2 and 2 gig ram
8800GT 3909 (177.35 driver)
SMP1 1560 (p2665)
SMP2 1634 (p2665)
total ppd 7103

I still have affinity changer loaded on this boxen. The GPU client takes 10-20% cpu, but I think it might want more. Without the SMPs running it had been getting up to 4400+ppd according to FahSpy (this boxen runs off a wireless connection and for what ever reason FahMon 2.3.3svn usually, but not always, lists this client as hung. I do have the GPU client set to "slightly higher" priority.

 
Sunin said:
I just d/l'd seesaw pro.. very cool tool, but I'm curious is when the service comes back up after a turn-in will it reset the affinities?

Is there any tool that can perm lock a process to a set core?
Click to expand...

The only other program I know which will help set affinities is THG Task Assignment Manager.
Read about it and download it at Tom's Hardware.
I have not used it so don't know how usefull-easy it is.

Luck ............. :D
 
Vista 32bit sp1 on X3350 @3.56ghz
running a single winSMP
8800GT at 600/1780/900 on 177.39 drivers

That's netting me about 2400 on the SMP unit and just under 5k on the 8800GT for about 7400PPD. I'm not bothering with the VM's or dual winSMP's, they're a little more hassle than I deemed worth on my only home pc.



 
WINXP 32 OS 2GB RAM
Q6600@3.5ghz 1smp client
8800GTS 512MB GPU client
1 Console client, 400ppd is on the low side. With the right WU can reach as high as 1000ppd


SMP 2300 + 5700 GPU + 400 Console = 8400 ppd


 
HighYield said:
WinXP32 Q6600@3.2 and 2 gig ram
8800GT 3909 (177.35 driver)
SMP1 1560 (p2665)
SMP2 1634 (p2665)
total ppd 7103

I still have affinity changer loaded on this boxen. The GPU client takes 10-20% cpu, but I think it might want more. Without the SMPs running it had been getting up to 4400+ppd according to FahSpy (this boxen runs off a wireless connection and for what ever reason FahMon 2.3.3svn usually, but not always, lists this client as hung. I do have the GPU client set to "slightly higher" priority.

Click to expand...

The above was with GPU core version 1.05
The GPU core has been updated to 1.06
Now the same 8800GT is getting 2155 ppd
That is a big drop in production. Anybody know what is up with the new core?
 
Play with the affinity or reboot. Normally you shouldn't lose more than 500 ppd. My 8800GT lost 300 ppd only while the 8800GTS 640MB lost close to 500.

Anyway, Vijay Pande now force core 1.06 on everyone so even if you are trying to hold on it, it will not work. It include a very important science fix (there is a problem with temperature control which cause the results to be somewhat inaccurate and 1.06 fix this to make more accurate results).

 
Air Cooled 420 said:
WINXP 32 OS 2GB RAM
Q6600@3.5ghz 1smp client
8800GTS 512MB GPU client
1 Console client, 400ppd is on the low side. With the right WU can reach as high as 1000ppd

SMP 2300 + 5700 GPU + 400 Console = 8400 ppd
Click to expand...
Which OS are you using and is the SMP client in a VM? I can't see all of these clients being run without conflict natively.

HighYield said:
The above was with GPU core version 1.05
The GPU core has been updated to 1.06
Now the same 8800GT is getting 2155 ppd
That is a big drop in production. Anybody know what is up with the new core?
Click to expand...
The new core is responsible for the drop in performance. Some have experienced a major reduction like you while others didn't and a small percentage claim improvements. It seems to affect XP users more than Vista since most of the Vista users reported little or no changes in production. The new core permits better quality science and reduces sluggishness in video apps. As I mentioned in another thread, it sucks but this is the shape of things to come. Try OC the card to reclaim some performance if you haven't done so already.

 
Originally Posted by HighYield

WinXP32 Q6600@3.2 and 2 gig ram
8800GT 3909 (177.35 driver)
SMP1 1560 (p2665)
SMP2 1634 (p2665)
total ppd 7103
affinity changer is loaded on this boxen.

Ok, seems like the GPU client was starved for CPU time. When I quit the SMPs and restarted the GPU client it was pulling ~4400-4700 ppd (back up from the very low 2k ppd). I loaded on the beta 8 gpu client and fired up the 2 SMPs and I was getting about the same 7k ppd I had started with. Seems like the GPU client is still not getting enough CPU time ~10-20% where it probably wants more (since its ppd dropped from 4400-4700 by itself to ~3700 with the SMPs running). 64 bit vista doesn't have this problem. I set the gpu client to normal priority and set it to run on all of the 4 cores but the %cpu utilization didn not increase. For folks running Winxp 32 bit GPU client and SMP(s) is the recommended procedure still to run affinity changer or will that mess things up now that we have the gpu client in the mix too.

Edit. oops - I missed Sunin's post that he was running 32 bit WinXp. No affinity changer it is then.
 
So some of you guys are running two instances of a VM with SMP in it, one GPU2 client, and a Console client all on one Quad core CPU?

And from what I can tell, it looks like Vista is the best place to run the GPU2 client since it uses up so much less of the CPU.

Am I correct in both those statements?
 
lelliott731 said:
So some of you guys are running two instances of a VM with SMP in it, one GPU2 client, and a Console client all on one Quad core CPU?

And from what I can tell, it looks like Vista is the best place to run the GPU2 client since it uses up so much less of the CPU.

Am I correct in both those statements?
Click to expand...

Yes, you are correct :)

 
SMP with VM seems like a lot of extra messing around to me... I'm running 2x SMP and 1x nVidia GPU clients, with Affinity changer on Vista32, 4 gigs. Currently getting 7700PPD on 8800GT OC, Q6600 @ 3.4GHz. Really depends on SMP project, 2665 is rarely above 1400PPD, but 2663 is closer to 2k PPD.

It does appear that Affinity Changer IS changing the GPU client to a single core... whether that makes a significant difference or not I'm not sure yet.

Seems like GPU is only using 2-5% of a single core anyway, so affinity probably doesn't matter much if at all.
 
Ok. Stupid question. Does Affinity Changer affect GPU Client? What happens if we use the new core and tell it to spread across all 4 cores - will Affinity Changer disregard that and assign to one core?


 
capreppy said:
Ok. Stupid question. Does Affinity Changer affect GPU Client? What happens if we use the new core and tell it to spread across all 4 cores - will Affinity Changer disregard that and assign to one core?


Click to expand...

This is absolutely not a stupid question. I loaded on the new core last night and spread the GPU load over all 4 cores and then 2 cores. The GPU was getting plenty of CPU power to work close to max production, however I run 2 instances of SMP on this WinXP 32bit Q6600@3.2 box. Normally I run with affinity changer but I took it off as the new core was supposed to be better. It turned out that my GPU ran fine but overall my system lost about 800-1000 ppd based on the SMP production not being all that efficient. Just before work this morning I reinstalled affinity changer and it looked like the SMPs were picking up production at a little bit of the expense of the GPU (the gpu core still had the affinity set to the 2 cores I had assigned). (May need to spread the GPU load over 3 or 4 CPUs). I'll report the results tonight when I get home from work.

Sunin- how are your boxen running now with the new core and no affinity changer?
 
Update: after running all day with the GPU2 client (v. 1.07 ) using 2 cores it still wasn't getting enough CPU power. I changed the affinity for the GPU client to use all 4 cores, and kept affinity changer going and this returned the boxen pretty much to normal production (posted earlier in this thread). For me with WinXP (32 bit) and 2 instances of SMP along with the GPU client it is required to keep the affinity changer going. If anyone can figure out how to do with out it (without going into VMs) I'd be interested in learning more about it.

 
HighYield said:
Update: after running all day with the GPU2 client (v. 1.07 ) using 2 cores it still wasn't getting enough CPU power. I changed the affinity for the GPU client to use all 4 cores, and kept affinity changer going and this returned the boxen pretty much to normal production (posted earlier in this thread). For me with WinXP (32 bit) and 2 instances of SMP along with the GPU client it is required to keep the affinity changer going.
Click to expand...
So, what you're saying is you have two SMP clients running with AC and one GPU2 client on all 4 cores of your quad, and you're seeing maximum output on your machine? How can this be when in Win XP the GPU2 client consumes an entire core and having no affinity set drastically reduces performance of the client?
 
I don't think it is maximum output, I would need to be running vista for maximum output out of the graphics card and whole system. I was saying it was more or less normal output for the SMPs (or you could read it the best I could get). The GPU is probably taking a bit of a hit. Today it has 2x p2665s @ 3621ppd and the 8800GT has a p5214 which is only giving it 3829ppd (other WUs with this set up have given 4.4k+ ppd....this may be one of the lower point ones or my setup might have drifted into some trouble) Total output today is 7450ppd which isn't too bad for XP and a Q6600@3.2. I haven't seen the gpu client take the full core in XP (and maybe that is some of my problem) but it does take a hell of a lot more than it does in vista. When I was testing this system it seemed to me that if I set the GPU client to just one core it wasn't getting enough cpu time. I also did set the GPU client to a higher priority over the SMPs...if that actually helps any.
 
Today I finally had time to do some serious experimenting:

Results:

GPU & 2xSMP Q6600 @ 3.2 Ghz

SMP1 741
SMP2 741
GPU 4365
Total 5,847

Remove 1 SMP client
SMP1 1974.86
GPU 4507.83
Total 6,482

GAIN over 600 PPD

I did this again on two other boxes and always netted between 600-900PPD. The GPU seems to get a bit more time and the SMP client seems happier.

On the same note 3.6Ghz seems to be the magic zone! Those seem to hold up well with dual SMPs and a GPU.

SMP1 1526.11
SMP2 1382.40
GPU 4365.47

Reducing to a single SMP would not net 2900 PPD.

Hope this helps others.


 
I find it weird that your 3.2 GHz Q6600 get about 750 ppd per SMP instance. Here's a sample from a boxen running at 3.3 GHz with 8800GT 512MB :

SMP1 : 1680
SMP2 : 2236
GPU : 4411

For a grand total of 8327 PPD.

My main PC with a Q9450 at 3.5 GHz with GTX 260 get this :

SMP1 : 2520 ppd
SMP2 : 2505 ppd
GPU : 6098 ppd

For a total of 11123 ppd

All stats taken from Fahmon and the SMP were all VMWare instances, not WinSMP.

 
Q6600@3.2 winxp 32bit, 177.35 drivers, 1.07 client , affinity changer active
SMP1 1547 (p2665)
SMP2 1656 (p2665)
8800GT 4000 (sometimes it's up to 4.4k) (p5217)
Total =7203 ppd

I am totally confused by this GPU client. I think I am running just about the same system that Sunin is, yet my SMPs are pulling in way more ppd at the slight expense of my GPU although the GPU does peak out at 4.4k ppd on occasion. :confused:

 
I did more messing around with my main machine since I put Vista64 on it.

Vista64
Q6600@3.6
8800GT w/ shaders@1805

Running VMWare Server 1.06 with two instances of stripped down openSUSE 10.3. Both VMs running 2605s at the moment and getting around 2200PPD each. Both VM instances are set to idle priority via Windows Task Manager instead of normal priority. (This works for me since I rarely reboot my boxen.) The GPU client is set to siphon off any spare cycles and uses 3-6% CPU and does around 5k PPD. This is a total of around 9400PPD for my main box. I don't have any other regular console clients running because they like to steal cycles the GPU client needs and at most I would get about 40-50PPD out of the regular console client.

fahmonnh9.png


I would be doing a lot better but one of my quads running Linux with dual Linux VMs are stuck with 2665s which has my total down about 800PPD.

 
HighYield said:
Q6600@3.2 winxp 32bit, 177.35 drivers, 1.07 client , affinity changer active
SMP1 1547 (p2665)
SMP2 1656 (p2665)
8800GT 4000 (sometimes it's up to 4.4k) (p5217)
Total =7203 ppd

I am totally confused by this GPU client. I think I am running just about the same system that Sunin is, yet my SMPs are pulling in way more ppd at the slight expense of my GPU although the GPU does peak out at 4.4k ppd on occasion. :confused:

Click to expand...

Its a trade off, you can set things up to favor the SMPs but teh GPU will suffer. most of my GPU's are at 4.4-5.4k. Depends on what affinities you are using. I'm going to test Xilikon's suggestion of 0 and 2 cores.


 
I found the best way to maximize PPD of both the GPU and SMP clients is to use Vista in a setup like mine. Two Linux VMs will use the vast majority of cycles to keep the SMP clients fed while the leftover cycles go to the GPU client (which doesn't hardly need anything) and lets it max out the PPD for that client. The only thing you have to do is drop the priority for both the VMs down to the lowest priority. They will still suck up almost all CPU cycles but allow the GPU client to grab as many as it needs.

This probably isn't as easy to work with on XP due to the GPU client needing many more CPU cycles to keep the video card fed. However, the same process would probably still work rather well except that the VMs won't get as many cycles as they would with Vista. The main advantage to both Vista and XP with using the VMs is that you will be more likely to get the 1760 point work units which crunch much faster than the 2665s. However, you can still get the 2665s on "dual core" machines or VMs.

 
You guys a re killing me - you give up as more points as I earn per day.... holy shmoley!

So if I follow this thread correctly and want to build a great boxen, I would want:

quad core (q6600 seems to be favored CPU)
4gb ram (can't decipher which speed is 'sufficent')
GTX 280 (or two if I can refinance the house)
Vista64
2 VM sessions of (32 or 64bit??) Linux for SMP folding with VM set to lowest affinity
GPU folding
HD does not seem significant, correct?
=====================
I can then get in the 6-8000ppd range with some tweaking?
 
oraldlight said:
You guys a re killing me - you give up as more points as I earn per day.... holy shmoley!

So if I follow this thread correctly and want to build a great boxen, I would want:

quad core (q6600 seems to be favored CPU)
4gb ram (can't decipher which speed is 'sufficent')
GTX 280 (or two if I can refinance the house)
Vista64
2 VM sessions of (32 or 64bit??) Linux for SMP folding with VM set to lowest affinity
GPU folding
HD does not seem significant, correct?
=====================
I can then get in the 6-8000ppd range with some tweaking?
Click to expand...
Easy, but I'd go with a GTX260 the money you save you can put in the 2nd 260. ;)

My 260 is pushing 6k points on it's own. The SMP will get you above 8k. Only do the 280 if you need/want it for games, but IMHO the 280 is overpriced (even now with price cuts) for the gain you get over the 260.

Edit my 260 is doing 6600-6800 ppd depending on WU.

 
Kendrak said:
Easy, but I'd go with a GTX260 the money you save you can put in the 2nd 260. ;)

My 260 is pushing 6k points on it's own. The SMP will get you above 8k. Only do the 280 if you need/want it for games, but IMHO the 280 is overpriced (even now with price cuts) for the gain you get over the 260.

Click to expand...

I agree with Kendrak here. The 280 is great for gaming, but the price premium for folding probably isn't worth it unless you can easily afford it. Most of us here are wondering about how to pay the electric bill so you need to be frugal at some point.

I have my GTX260 cranking out at ~6700 ppd (621/1512/999). The 260 is even a bit of a premium for a pure folding box as you could get more production cheaper from 2 8800GTs.

 
HighYield said:
I have my GTX260 cranking out at ~6700 ppd (621/1512/999). The 260 is even a bit of a premium for a pure folding box as you could get more production cheaper from 2 8800GTs.

Click to expand...

The 2x 8800GT is a good option if:
*You have the slots to put them in
*You have a PSU that can support them (if you are going to run a 280/260 you probably already have one)
*You don't mind going deaf from the fans in those things (crazy loud after you OC the hell out of the card ;))

 
Interesting (and great) results all the way around!

I've tweaked and tweaked and come up with what I think is the best results attainable for the parts I have (currently ;)). First the specs:
Q6600 @ 3.05 (never breaks 52c and seems to be the sweet spot for summer fun)
slot 1: 9600GT @ 667/936/1884
slot 2: 8800Gs @ 615/800/1884

I've been following this thread on the FF and it seems to be pretty well documented that running different models of GPU in multi setup causes the second card to suffer about 10-20%+ loss of it's full single slot dedicated speed. My next attempt at tweaking will be to setup the 8800GS as the primary card and see if I can get that magic 10000PPD I'm almost at. When I started tweaking, I was running high 9.6k - to low 9.7kPPD.

Final (so far) results and method to achieve: note this won't work for a hands off setup. I have to get this setup manually, but since I don't shut down, unless I lose power, it's all good.

I start SMP Seesaw at boot, then fire off slot 1 GPU and set affinities for both executables to core 3. Then fire off slot 2 GPU and set both executables for cpu2. Next I fire up VMware and load NotFred diskless CD w/usb backup and set affinity to cpu's 0-1. All other processes are left on 0-3 I also set priority on both GPU cores to normal.

Slot 1 GPU: 4066-4079
Slot 2 GPU: 3759-3787
VM SMP: 2049-2057

Total: 9874 - 9923

The 8800GS should pull 4300-4800PPD so I still have to try that in slot one and see if I with the PPD I lose from the 9600GT if I can pickup the extra 80-125PPD I need to :).

The moral of the story is that if you're gonna run multi GPU, keep the cards the same in however many slots you fill....:D I wish I knew that going in... :confused: No worries.. it's all good.

 
You must log in or register to reply here.
Back
Top