F@H: 'Interesting' stuff

[Insidious]

So, those of us that chose to use -advmethods and give this core14 a whirl are seeing some new behaviors on our NV GPUs. Here's a couple observations at my end. Nothing broken or action needed. Just thought it might be a good time to 'share the goodness'.

First of all, my PC setups: I have two PCs which are identical in terms of hardware. (OC'd Q6600 with two video cards. Each PC folds one 8800GTS (512MB) and one 9800GTX+. I am also folding a single SMP WU on each PC alongside the GPUs.

I am using the 178.24 version drivers which let the GPU clients run a little faster at the expense of more CPU cycles required to support them (A lilttle less SMP PPD, but not as much as the gain in GPU PPD). I have set the Windows System Environment Variable 'NV_FAH_CPU_AFFINITY' to the value of 0 to allow the GPU clients to scavange cycles from any core they would like.

I think we've all seen the cooler overall temperature of our GPUs with the 59xx series work units. Those that have watched (GPU-Z or RivaTuner are what I use) see that the temperature fluctuates a few degrees every few seconds. My first observation that I thought was interesting is that the temperature fluctuations are not the same on the 8800GTS as compared to the 9800GTX+. The 9800GTX+ card has larger fluctuations which are also less frequent. So I assume, these temperature fluctuations will be a function of what card you are using.... This may also go a long way to explain (or at least correlate to) why the fluctuation in frame completion time is also a function of the partucilar GPU you are using. (As reported at the fah forums... )

The other thing I have seen is that the 59xx Work Units show more affect of CPU speed than I have noticed on the WUs we were crunching before. One of my PCs is clocked at 3.2HGz and the other is clocked at 3.12GHz. I wouldn't have thought this would make any significant difference in the GPU frame times, and it didn't before. But now with the 59xx WUs, I can actually see a consistant difference (fahmon). It appears (now that I've watched for a couple of days) that this small difference in CPU speed is leading to ~200PPD difference in the 8800GTS cards and ~300PPD difference in the 9800GTX cards. I haven't had time to mess with creating larger speed differences between the two PCs (particularly reversing which one is the faster one) to verify this.... but I will. This stuff has me curious.

Anyway, hope you enjoyed the read

-Sid

 

[Drsignguy]

Enjoyed the read Sid, Thanks. :thumbsup:

I would second on most of your information that you have provided. Only thing I have done differently is I have all my Os's with 182.06 version drivers and let the Gpu's do their thing and am trying out 2 Smp's on each quad core to see how well they react with little to no Gpu interference. So far they are doing pretty good.

Also, as mentioned in the "new WU" thread, The temps have dropped considerably and the only thing I did was drop the fan speeds back to "auto" and not a fixed speed, also to see how they react.

 

[Markfw]

I have 181.22 drivers, and 2 smp on a quad, one on a duo, and 11 systems total, 6 quads, 4 duos, and a celeron. 19 cards scattered on the boxes. I don't see much variation on cpu to ppd. I think its the per-card variation. I do have 6 identical 9800 GT's. On the 6 cards, it seems like cpu speed DOES change ~200 ppd or so on my Q9450. That might explain my 260 card, since its running on a 2.8 ghz celeron 420.

 

[Drsignguy]

In 2 of my rigs, they have Identical cards and of all of my rigs, the lower card seems to always have a lower PPD....

 

[Insidious]

Here's how I always understood the two vs. one SMP thing:

In windows XP, the OS does not do a economical division of the work for multi-threaded applilcations. So, even though a single SMP was designed to use all the free cycles of all 4 cores, it lost some due to 'overhead' (what ever that means)

When you ran two SMP work units on windowsXP, the 'overhead' stayed the same, but you had two clients running with an effective use of more cycles than with only one. Then, you add in the Affinity Changer to assist core assignment, you could see as much as a 20% increase in PPD with two SMP WUs and the Affinity Changer over a single SMP. (IF you are running only SMP clients.... no GPUs)

This difference was much less pronounced with windows Vista because of kernel differences not letting so many cycles go to waste.

Here's my take on why much less CPU speed dependence is observed in GPU PPD when the 18x.xx drivers are used:

The GPU client uses almost no CPU (difference more pronounced with WindowsXP than with Vista) with the 18x.xx drivers. Most users report a 15 to 20% decrease in a given GPU ppd when compared with 178.24 drivers. The upside is that you have more CPU available when you use the newer drivers.

On my PCs, the decrease in GPU PPD (remember the difference is doubled since I have two of them) due to the 18x.xx drivers was MUCH greater than the increase in SMP ppd due to the freed CPU cycles. So overall, I have significantly better PPD with the older drivers (faster GPU, slower SMP) than vice versa.

Also, when I compared the difference in PPD from two SMP clients, which slowed my GPUs because I use the older drivers (178.24) and the PPD from a single SMP.... I get nearly 1,000 better overall PPD with only a single SMP client running.

I highly recommend that if you are running WindowsXP on a quad and two video cards you experiment for yourselves. Try the older drivers (178.24) for the GPUs and set the environment variable as I mentioned in the OP. Then shut off one of your SMP clients.

I really do believe you will see the same PPD increase that I have.

-Sid

NOTE: All the above ONLY appies IF you are running WindowsXP on a QUAD core.

edit: If you are running a DUAL core processor and a single GPU on WindowsXP, using the 18x.xx drivers will free up enough CPU cycles to add an SMP to maximize your PPD. (Hard to get SMP finished in time if you run the 178.24 drivers on a dualie)

 

[Markfw]

Running 17x.xx and one smp and 2 gpu, I did OK< but when I went to 18x.xx and 2 smp, the total pdd was higher, so thats where I am at.

 

[Drsignguy]

Yes, I too just simplified and just stuck with the same driver ( for each card, for each OS ) and all is good. I just started up the (2) SMP's just the other day as I wanted to see what my max PPD could be and then all of a sudden Boom, these new WU's came into play and amped it up a bit more it seems.

 

[Markfw]

Here is my current farm.

Once card is down (bad) and one due to gaming, but you get the idea. All the ones with GPU in the name are the video cards. My total PPD varies a lot from (when all cards are running) 110-125k ppd.

 

[Drsignguy]

That's more like a ranch! Holy Hardware! Makes my little setup look like a petting zoo.


Here is my lastest Snapshot of all my rigs with the new client.


Nice Ranch Mark!

 

[Markfw]

so you have 3 quads, and one duo, 6 260 cards, and 2 8800 gts cards. Your cards are more powerfull than mine each.....

Pretty compact powerfull farm there !

 

[Drsignguy]

Thanks Mark

I have 1 more 260 55nm to get and then I will have all 5 of my rigs complete for what I want them to do. But yes as it seems, it is pretty compact and powerful.

 

[Duvie]

I went from 178.xx cuda drivers to 181.xx drivers and I only saw a drop of about 200+ppd....so nowhere near the 15-20%....

I liked it eventhough my ppd went down....

1) I dont run cpu client anymore so I keep my case temps down which can effect my gpu temps
2) with older drivers I would get more nv4 infinite loop BSOD....now I rarely see one unless I do something stupid like startup game while GPU_0 is going....


So you are saying this is taking cpu into account and that is why you are seeing an increase.....interesting....may have to test this out.....

 

[Insidious]

That sounds really encouraging that your drop in GPU performance was small !!!! I'm assuming that's the 9800GT. Are you running WinXP?

I didn't really understand question in your post

So you are saying this is taking cpu into account and that is why you are seeing an increase.....interesting....may have to test this out.....

If I remember right I saw ~16% drop, but it's been a while and it was on the first of the official 180.xx drivers. If the newer ones reduced that to the 5% you are seeing, I'm going to switch too!

-Sid

 

[Insidious]

YAHOOOO

I just tried the 182.08 drivers and I don't see ANY significant change in my PPD (GPU or SMP)

I'll have to watch it a while longer (only had about 15 minutes to watch while still on the same WUs) but it looks like I can go back to running latest and greatest drivers with no issue what so ever!

-Sid

(looks like there has been some good work done by NVidia between versions 180.60 and this one... 182.08)

edit: Here's a screenie comparing the fahmon reports with each driver (same WUs)

The GPUs are down (very slightly) and the SMP is up (very slightly) and it all ends up a wash for total PPD

 

[Markfw]

But what that DOES mean is that you can now run 2 smp clients and 2 gpu clients. About 4k more ppd.

 

[Insidious]

LOL

No, the one SMP WU is not only using half of the CPU. The only gains from running two SMPs instead of one is capturing the 'overhead' that is wasted.

The first time I turned you on to the affinity changer and two clients, I think I remember us seeing something like a 25% gain.

but yes, I will be trying it when I have time to install SMP #2... I'll let you know. You can't imagine how much I wish you are correct!

-Sid

 

[Markfw]

I have 6 boxes doing 2xsmp, and yes, under the new drivers it gives better total ppd.

 

[Drsignguy]

Originally posted by: Markfw900
I have 6 boxes doing 2xsmp, and yes, under the new drivers it gives better total ppd.

I must agree with this.

C'mon do it Sid, I know you want to.....:evil:

 

[Markfw]

BTW, I just finished installing the 182.08 drives on al 19 cards in 11 boxes (boy does that take some time) and I see no ppd change...

RATS

 

[Insidious]

Originally posted by: Drsignguy

Originally posted by: Markfw900
I have 6 boxes doing 2xsmp, and yes, under the new drivers it gives better total ppd.

I must agree with this.

C'mon do it Sid, I know you want to.....:evil:

:laugh:

They're up and running. I'll pop in before bed and add to that screenie to let you know how we did. (I forgot how fast it is to install the second instance. Had to read my own guide... LOL :thumbsup: )

 

[Markfw]

copy your folder for the first mp client, then delete queue.dat,client,.cfg and unitinfo.dat and the work directory. Make a shortcut to the exe, and run it. It will ask for config info, just make sure to set the machine ID to one that is not used by the other 2 clients.

You are done.

 

[Insidious]

:laugh:

Yes I am done...

But thanks anyway!

 

[Drsignguy]

Originally posted by: Markfw900
BTW, I just finished installing the 182.08 drives on al 19 cards in 11 boxes (boy does that take some time) and I see no ppd change...

RATS

Well removing 19 drivers and reinstalling the lastest, I would think it would take a bit of time. Did they all go smoothly or did you have some troubles??

Originally posted by: Insidious

Originally posted by: Drsignguy

Originally posted by: Markfw900
I have 6 boxes doing 2xsmp, and yes, under the new drivers it gives better total ppd.

I must agree with this.

C'mon do it Sid, I know you want to.....:evil:

:laugh:

They're up and running. I'll pop in before bed and add to that screenie to let you know how we did. (I forgot how fast it is to install the second instance. Had to read my own guide... LOL :thumbsup: )

Now that is funny. Been a while huh???? :laugh:

 

[Markfw]

smooth as glass.... Just takes time. And tomorrow 2 more 9800 GTX+ cards come (one to replace a bad one, and one more) 756 mhz core factory. Right here

So it will be 20. I am really running 18 now, due to the bad one.

 

[Drsignguy]

Those cards seem to be putting out some nice numbers. The price looks very reasonable for what PPD your getting out of them. Hmmmm, I will have to check these out in the near future. First up is my 7th 260, then I am looking into a few more cards like this...

20 is a nice round figure and a very respectable PPD production.