• 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.

Your GPU folding configuration

J

Jon855

[H]ard DCOTM January 2008
2FA
Joined
Aug 25, 2005
Messages
12,132
In this thread, please only share how you set up the GPU Folding / CPU Folding... This is not meant to be a discussion thread. This thread is meant to share with others how we have our folding rig set up. If there are any off topic post, I'll delete it. Please keep this clean and tidy for the others to take a look at our configurations. Thanks.

Try to keep it in the format of :

6.5Catalyst

AMD X2
Core1 - idle - 95% - MachID:1
Core2 - idle - 100% - MachID:2

X1900XTX
GPU - low - 100% - MachID3
 
omega drivers

opteron 170
core 0 at idle 95%
core 1 at idle 95%

x1900xt
low 100%


T-3
 
Just got done with a series of tests this evening and here is what I got:

System:
X2 3800+ @2.24Ghz
X1900XT 256MB (3D forced to 621Mhz GPU, 720Mhz Memory)
1024MB memory

The cores I had during this test were:


I ran 5 configurations:
P2414 for both CPUs, P2711 for the GPU

1. GPU Only

Setting: 100%, Idle

GPU: 362PPD - 8.76 Minutes to complete a frame

2. 2x CPU Only (My normal configuration)

Setting: CPU1 & CPU2 100%, low

CPU1: 257PPD - 33.62 Minutes to complete a frame
CPU2: 257PPD - 33.62 Minutes to complete a frame

Total: 514PPD

3. 1x CPU & 1x GPU (recommended Stanford setting)

Setting: GPU 100%, Idle; CPU 100%, low

GPU: 359PPD - 8.83 Minutes to complete a frame
CPU: 267PPD - 32.38 Minutes to complete a frame

Total: 626PPD

4. 2x CPU & 1x GPU (1)

Setting: GPU 100%, Idle; CPU1 100%, low; CPU2 95%, low

GPU: 67PPD, 47.28 Minutes to complete a frame
CPU1: Had not completed after 1.5 hours
CPU2: 249PPD, 34.72 Minutes to complete a frame

Total: Less than 400PPD (unstable, would not recommend)

5. 2x CPU & 1xGPU (2)

Setting: GPU 100%, Low; CPU1 100%, Idle; CPU2 95%, Idle

GPU: 357PPD - 8.88 Minutes to complete a frame
CPU1: 145PPD - 59.43 Minutes to complete a frame
CPU2: 142PPD - 60.77 Minutes to complete a frame

Total: 644PPD

So the results from this first set of tests is that the recommended GPU solution provides the best PPD, at least for the cores involved in my system and with my current system specs. 600 point WUs are about the second best core I can receive.

From a power standpoint. It makes more sense to just have 2 CPUs running (Point per Day per Watt). I'll get Kill-a-watt readings later.

I will start tweaking the percentages to see if there is a better solution that causes less conflict between the cores.
 
Mike Houston says the X1900s are pulling about 80 watts, not nearly the full 3d loaded power draw of around 125w. It's like adding the power draw of another CPU. ;)
 
Yep, people are posting Kill-a-watt reading that show in this ballpark, although it's hard to separate from the entire system as we do kick the CPU up a little as well. And of course there is PCIe interactions and memory. ATI's lab measurement was right around 80W for the board.

We shall see how this plays out. Perf/Watt, it's a no brainer that the GPU dominates here. PPD/Watt is more complex as it depends on the CPU comparison. PPD/$ would also be interesting for someone to take a look at. Of course, there is already much discussion/commentary about the GPU PPD calculation currently used, and no, I'm not going to jump in. ;-)
 
Mike, its a real treat to have you on our boards here at [H]ardOCP.. However, you would look less out-of-place without the 'noobie' beside your name above all of your posts..maybe someone can change that for you, something like 'honorary [H] gawd' lol.. thanks for all the great info and we appreciate everything you're doing, it truely makes a difference in this community.
 
On other forums, you can change your badge, but I can't figure out how to do it here. I'd settle for 'GPGPU guy'. ;-)
 
One big difference for me is that my typical configuration is with an IGP so the base wattage of the system is much lower. I think that system was running about 128 watts before installing the X1900XT. This puts the PPD PW at about 4.

To get the same PPDPW using the X1900XT (626PPD). I would only be allowed to be at 157 watts. I guessing that I am over that. Once I get done with the next series of tests I will put the kill-a-watt in line.
 
mhouston said:
On other forums, you can change your badge, but I can't figure out how to do it here. I'd settle for 'GPGPU guy'. ;-)
Click to expand...


Either post a lot or insult Kyle ;) Only way to change it here (or be a mod....)
 
Mike,

Have you explained somewhere why the FAH GPU client conflicts so much with a CPU client when on the same core? They really don't like each other.

The other interesting thing is that IDLE CPU loses to Low GPU. This is counterintuitive. (I've been playing with affinities....
 
Part of this is the limitations of DX9 which are supposed to be addressed in DX10. The main issue is that we need to keep the board fed with commands and data, and there isn't a good asyncrhonous path for this. Basically, and simplified, we hand a program and data off to the GPU, then have to poll for completion, and then give it more to do. The kernel and driver do the polling for completion. Notice in taskman that kernel times count for much of the GPU load. With DX10, the overhead of this is supposed to be better. In something like CTM, we can be much smoother about this and do lighter weight checking of the hardware because we have more direct control.

There are also a few things we can do to try to improve this, but they involve pretty hacky things things. This is something we'll look at once we stop seeing EUE's and other annoying but less "oh crap" issue with the client like the completion printout's disappearing.
 
Addendum: when you play with priorities, you are effectively doing what a sleep inside the code would do. Effectively, you are shrinking the time slice given to an app and yeilding more time to others, but in a heavy-handed kernel level kinda way.

Some people have had success setting priorities and CPU% on the CPU client to still allow the GPU client to run well, maybe with a small hit, and also keep the CPU clients chugging okay, but maybe a little slower.

I thought there was a forum post here on this...
 
Cool. I am aware of the percentages and loads. I just wanted to control which CPU client the GPU client was fighting with. Best way to ensure this is through setting affinities. At the moment I am lower the GPU % to give some time to the CPU client on the same core.
 
As one without an appropriate GPU, I'm enjoying the vicarious folding satisfaction
from those team members who are tweaking and experimenting with the GPU Beta.
Much appreciated and keep the information coming. :)

It would be interesting, during this Beta period, to have one of you switch the
username of your GPU console to GPU_x1900, or something
unique and descriptive, so that we can monitor what you are getting from the
GPU "alone." Of course, after a week or so you could change the name back
to Celerator...I mean to your regular name... ;) ;)

Fold on [H]orde!

 
I have to admit I haven't finished a WU yet. At 47 frames currently. I've had a few Early Unit Ends.


I am trying: CPU1 95%, CPU2 95% (Idle) and GPU 100% (low) at the moment. It looks like the GPU is taking 1 core and the 2 CPU clients are sharing the other core: CPU1 25%, CPU2 25%, GPU 50% on the process load. It's probably like running 2 CPU clients on 1 core (about 5% gain). I haven't completed a series of WUs yet to verify.

 
mhouston,

How much of the core and client are using CTM right now? If nothing, will CTM provide the answer to the conflicts with drivers and Windows? I assume with CTM that it'd be a direct communication and no use of card drivers would be needed, or am I wrong?

 
CTM is too new, so everything is currently running on DX9 using Brook. Once we can do a Brook port to CTM, then we should be able to get the current client up, but CTM needs to be publicly available before we can do that. We could also bypass Brook and to straight to CTM, but that will likely take more developement time. CTM would allow us to avoid driver breakage. What generally breaks GPGPU apps is the compiler in the driver, and this is what we hit in 6.8/6.9/6.10 drivers. Since we compile offline and load an binary with CTM, we can avoid the driver breaking us once we get things working. More importantly into the future is that CTM will allow us to do things we couldn't in the past with extended functionality.

CTM is not going to fix the issues with trying to run other graphics apps on the GPU at the same time.
 
Hito Bahadur said:
I have to admit I haven't finished a WU yet. At 47 frames currently. I've had a few Early Unit Ends.
Click to expand...

The first machine I put my X1950XTX in gave me nothing but EUEs. Now that I moved the card to a different machine with a little better power supply that's running WinXP instead of Win2K I'm on my 5th WU with no EUEs. During the switch Pandegroup came out with a new 0.06 core also, so I don't have a clue which one of these changes cured my problems.

This is my setup:

Asus A8N-SLI Premium
AMD X2 4800+
Corsair TWINX 2048-3200C2 Pro (2-3-2-6-1T)
Powercolor X1950XTX
Turbo-Cool 510 Express / SLI
WinXP SP2 fully updated
nForce4 6.86 chipset drivers
ATI 6.10 Beta Catalyst
ATITool 0.25 Beta16 (used to force 3Dclocks)

Note: When you get to 85 frames and it appears the client has stopped, don't panic. All 4 of the WUs I've done so far stopped showing progress after 85 frames in the log, but the client is infact still processing the WU and all 4 WUs finished.


Hito Bahadur said:
I am trying: CPU1 95%, CPU2 95% (Idle) and GPU 100% (low) at the moment. It looks like the GPU is taking 1 core and the 2 CPU clients are sharing the other core: CPU1 25%, CPU2 25%, GPU 50% on the process load. It's probably like running 2 CPU clients on 1 core (about 5% gain). I haven't completed a series of WUs yet to verify.
Click to expand...

This is the same setup I'm running with right now. It seems to be working well.

Be aware that task manager is giving false info as to what CPU resources your GPU client is using. The polling that Mike mentioned makes it look like the GPU client is taking all of one of your CPU cores, but it isn't.

ChelseaOilman

 
I noticed a problem with the log also, but only when running 2xCPU and 1xGPU:

Specifically, I had an 1809 protein runnig on 1 core, a 2409 protein on a second CPU and a 2711 on the graphical client. The 2409 FAHLOG never updated until I rebooted and then it showed that frames had been complete.


I also ran into a stability issue with the setup above. Specifically my system has froze after about 3 hours twice.

I went to 1xCPU and 1xGPU since it doesn't seem to be garnering a significantly greater number of PPD and I have the stability issue with that configuration at the moment (might look into it later, plus I need to switch to a better PSU just in case).

I've moved on to power calcs:

I'm running the following system:
Winfast S939 6150 mobo
X2 3800+ @2.18Ghz (218FSB)
2x512MB Corsair Value select PC3200
2x80GB HDD (seagate)
1xGeneric CDROM
350w Antec PSU
Sapphire X1900XT 256MB at 621CPU (1.4v) and 720MEM(2.05v)
Stock Heatsink for CPU and Video


This system is drawing an average of 225watts at the wall and with the 1xCPU (P2409) and 1xGPU (p2711) is producing 629PPD

Without the X190XT and running 2xP2409 and just using the IGP the system draws 128 watts while producing 526 PPD.

This equates to:
1xCPU/1xGPU - 2.8 PPDPW (points per day per watt)
2xCPU/0xGPU - 4.1 PPDPW

or financially ($0.135 cents per kilowatt here in RI)
1xCPU/1xGPU - $1.16/KP (per thousand points)
2xCPU/0xGPU - $0.79/KP

or from a budget standpoint:
1xCPU/1xGPU - $0.728 per day ($21.83 per month)
2xCPU/0xGPU - $0.415 pe day ($12.44 per month)


Pretty hefty extra cost for running GPU versus CPU

From a power cost standpoint (for my case) an additional 271 points need to be made up somewhere.

Either the GPU points will need to be worth more to make up for not being able to run the second core or the second core will need to be able to run another instance of FAH CPU client.

The right solution is probably a combination of the two, with the other thing being the need for the GPU companies to start producing more efficient video cards.

Comments?
 
Of course which proteins are foldig does change the numbers. A more GPU friendly instance would be using theP1809 core (much lower PPD for AMD processors).

CPU = p1809 core
GPU = p2711 core

1xCPU, 1xGPU produces about 557 PPD
2xCPU, 0xGPU produces about 402 PPD


1xCPU, 1xGPU = 2.48 PPDPW
2xCPU, 0xGPU = 3.14 PPDPW

In this case, 150 points need to be made up somewhere to make things equitable for a power consumption standpoint.


For Core 2 Duo's there is no advantage to using the GPU client as the PPD from the 1 lost Core is greater than the GPU can produce while still incurring an 80 - 100 watt penalty (100 watt if you would use an AGP otherwise).

These numbers don't address the additional $250 up front cost.


I like the GPU, but for those of us who pay multi-hundred dollar power bills because of folding it isn't making financial sense at the moment.
 
Here are some Kill-A-Watt readings for my setup running with 3D clocks 650/1000.

Computer @ idle = 195 watts
With F@H CPU client processing a p2126 WU = 243 watts
With CPU client processing a p2126 WU and GPU console client processing a p2725 WU = 338 watts

Note: The watts used by the monitor is not included in those readings.

I took those readings yesterday. Now I'm running a second CPU client and the total wattage right now is 340 watts.

CPU client #1 = p1495
CPU client #2 = p2409
Gpu client = p2725

ChelseaOilman

 
Is that a X1950? Assuming so because the mem clock is so high. It seems that those draw a little bit more power on the memory side than the X1900s. GDDR4 vs. GDDR3
 
Here's the break even equation (based on 80watt difference between 2xCPU and 1xCPU + 1xGPU)

GPU PPD needed to break even from power requirements = t*(160/y + 1)

Where:
t = PPD for 1xCPU
y = watts of system running 2xCPU core

for example:
A system running at 240 watts and producing 260 PPD for 1 CPU client instance would need to produce the following PPD to make up the extra 80 watts of powering the GPU.

t = 260 PPD
y = 240 watts

260 * (160 / 240 + 1) = 433

In other words the GPU must produce 433 PPD or bette to make sense running GPU from a power standpoint

Interesting eh? If someone can check my math I would appreciate it.

edit: Corrected info.
 
More detail:

y = watts when running 2xCPU
t = PPD for 1xCPU
z = watts when running 1xCPU + 1xGPU
k = PPD for 1xGPU

2*t/y = PPDPW (Points per day per watt) for a 2xCPU setup

2*t/y * z = PPD needed to have same PPDPW as a 2xCPU setup

k + t = PPD from running 1xCPU + 1xGPU

2*t/y*z - (k+t) = deficit or surplus from running 1xCPU + 1xGPU

Setting equation to zero and solving for k should give the required PPD from the GPU to break even.

k = 2*(t/y)*z - t

Simplifying equation for 80 watt difference:
z = y + 80

k = t*(160/y + 1)

edit: Simplified equation was wrong. Right now.
 
They should base the benchmark baseline of GPU folding points on that basic concept Hito.. give it enough points that it makes sense from a points-per-watt perspective, and then a little more to make up for all the extra work a GPU client does.
 
I'm kinda agreed. I think there will probably need to be adjustments, especially if they can't reduce the CPU load created by GPU folding. In essence it is taking the load of a CPU core and the GPU to enact GPU folding plus all the extra power requirement.

My current setup 1xGPU + 1xCPU with a P2711 and p2409 core respectively has a 308PPD deficit based on the equation above:

t = 264 PPD
k(actual) = 356 PPD
y = 128 watts
z = 225 watts

k(required) = 2*(264/128)*225 - 264 = 664 PPD

k(required) - k(actual) = 664PPD - 356PPD

deficit = 308PPD
 
Or, in my case. The GPU factor would need to be over 7x instead of 4x to make sense using a GPU to fold instead of CPU.
 
Hito Bahadur said:
Or, in my case. The GPU factor would need to be over 7x instead of 4x to make sense using a GPU to fold instead of CPU.
Click to expand...
That's pointwise but that's not the goal of F@H anyways. In reality they don't care about points at all, they just want results and you should agree to this regardless now fold away :eek:
 
I hear what you are saying Jon, but points do make a difference. In my case, if it's a choice between getting 2000 points per day using 1000 watts or 2000 points per day using 1400 watts, I'll say screw the GPU and just do CPU folding. It's my $$, my choice.
 
Hito Bahadur said:
I hear what you are saying Jon, but points do make a difference. In my case, if it's a choice between getting 2000 points per day using 1000 watts or 2000 points per day using 1400 watts, I'll say screw the GPU and just do CPU folding. It's my $$, my choice.
Click to expand...
Points have no bearing for the entire project at all. But if you're a whining point bastard... Go to Mary03 and cuddle with her. :D

 
Jon855 said:
Points have no bearing for the entire project at all. But if you're a whining point bastard... Go to Mary03 and cuddle with her. :D
Click to expand...

Points have no bearing, ability to return WU's are the objective.
 
Jon855 said:
Points have no bearing for the entire project at all. But if you're a whining point bastard... Go to Mary03 and cuddle with her. :D

Click to expand...
Let's be serious. While I am irrational in my behavior and will fold regardless, I think that the incentives should be set to make people indifferent between CPU and GPU folding from a "watt" perspective. This way, people have the choice to consolidate towards more GPUs or just keep the rigs that they have.
 
You don't understand my point Jon and I don't have the patience to explain it to you.
 
Hito Bahadur said:
You don't understand my point Jon and I don't have the patience to explain it to you.
Click to expand...

Chill.....he's a Noobie big mouth and ass-spire's to be like Marty, without
the credentials :rolleyes:
 
That's why I didn't waste time personally attacking him for being an ass. I just told him I didn't feel like explaining my point.
 
Hito Bahadur said:
You don't understand my point Jon and I don't have the patience to explain it to you.
Click to expand...
No I understand your point... But I don't see the points being a thing thant you should be concerned about after all you are not being rewarded with something if you win the most points.

I understand that you perfer to earn more points per watt to make it worthwhile "pointwise" but I'm arguring against that in non-pointwise.

Don't get me wrong I understand where you are going I just don't see why people make big deal over this pointwise when u are doing 30x the work over a cpu and that's what we should be happy doing regardless of the points. That's my only point here and no I am not going to say that your idea of requesting more points is bad or stupid elsewise, I just don't care is all.

Fold On....

zim01 said:
Chill.....he's a Noobie big mouth and ass-spire's to be like Marty, without
the credentials :rolleyes:
Click to expand...
I'm not that n00bie running my mouth all over the place... Who says I need credentials to do that? If so then that's the stupidest thing ever...

 
My point is if Stanford people think that 30x the work is being done then some valuation in the point system should reflect that. I'll do the most work I can afford and that will be based on points. It is up to Stanford to balance the points based on the best effect for science.
 
Hito Bahadur said:
My point is if Stanford people think that 30x the work is being done then some valuation in the point system should reflect that. I'll do the most work I can afford and that will be based on points. It is up to Stanford to balance the points based on the best effect for science.
Click to expand...
+1 I'll agree with you on that.
 
You must log in or register to reply here.
Back
Top