Post your Graphical Representation, of an App's load on the system.

I am getting interested lately, at taking a look on how an app's load gets balanced on my system. With GPU-z now able to show me each individual GPU's utilization, and MS's Performance Monitor can keep an eye the 4 cores of my Q6600, I think I'm getting a decent picture of my Total System Activity now.

We don't need to post our CPU or GPU clocks, and the score we received. That is not the point of this thread.

What is for me, is the quest for the app with the best load balance across my entire system using both CPU's, and GPU's.

For me, with running a dedicated PhysX processor, my balanced app would need to be CUDA, PhysX, or Direct Compute related.

Not so for everybody... That's fine too.

We also might get a few talking points, from what the utilization graph shows on various apps.
The rather alarming thing to me, is how few apps actually tap into all of your CPU's, and GPU's calculating abilities...

So I want to see if I can find one that will, or does a darn good job at using most available resources on the system.
It might be a little twisted I know, but I'm a sucker for apps with outstanding load distribution!
It ensures you are getting the best performance on your system that's possible, on the hardware installed.
(All hardware.)

What I'm looking for: I would like to see you fire up as many instances of GPU-z, as GPU's installed in your system.
Insure to use GPU-z's drop down selector, to point each instance at a different GPU.
Then, select the 'Sensors' tab, to allow us to see the GPU's utilization.
(You will need two instances of GPU-z running, to monitor a 295.)

Also start MS's Performance Monitor, to keep an eye on your CPU's utilization...

Let all graphs flat line like this, before you start your test:


You don't need to post your 'Flat Line' shot, I'm just giving an example.
I am only trying to stress, that we want to capture the app's processing activity, but nothing extra... 


So here I go...

System: (1) Q6600, (1) 295 in SLI Mode, and (1) 280 operating in dedicated PhysX mode.
App run was: The Vantage 'Crash-n-Burn' PhysX Test, Extreme preset.



Things that stand out to me is no use at all on my 295, and the PhysX test running on my 280, appear to never have exceeded 50% utilization...

I would have expected to have some minimum activity on my 295, and 90% or more load on my 280.

It does look like all 4 CPU cores made it to the party, and are well balanced, and utilized.

I will post mine for folding once this install is done.

Edit:
Here is folding on my 260 C 216.

Oh no,



My brain,




I...




Im...




Im THINKING! WOAH!

Thanks... It is a good app!

In my above posted utilization shot, I had hit the '-' key 30 times.
Each time the systems workload increased a micro-jiffie...  (That's why it might have looked a tad high.)

"QUOTE (cbuchner1 @ Dec 8 2009, 04:09 PM)

The reason it takes more time to render when you press "-" a lot is that you enter the range of negative exponents. Then I use a slower rendering method that uses more trigonometry (sine/cosine expressions and such). The fractals that result are of "mandelier" type (hollow objects that ressemble chandeliers)".

I did that, just to see when the systems workload increased, how well my (3) GPU's would split the load.

But yes, Mandelbub did indeed make excellent use of the PhysX card. Your getting 60FPS with the default screen, and I'm getting 90... The 280 is adding some performance for sure. 

I am like you, and wonder why our GPU's don't run at 100% more often...
That Vantage 'Crash-n-Brun' test I posted, was pitiful as far as GPU utilization goes.

cbuchner1 also posted: "The Optix libraries are good at distributing the work load among several GT2xx GPUs."

Which makes me think about the term 'Dedicated PhysX Processor'.

It's almost time we have a shocker thread started entitled: 'The Death of the Decicated PhysX Processor!'
Then you get inside, and figure out it's really all about CUDA, OpenCL, Compute Shader, and PhysX apps, and how more are processing on your 'Dedicated PhysX GPU' every day.

The term doesn't really fit so much any more...
It seems to imply - PhysX Only... (That would be incorrect!)
It needs an image change.

I vote: 'The Dedicated OMNI-Processor', or maybe 'The Dedicated Visual Computing Processor'.  
It's getting to be a more accurate description.

OK, back to load distribution!

I did reboot fresh, and only ran the Performance Monitor, and Mandelbub.
This is just Mandelbub's default screen's utilization. (I get 93FPS.)

For trivia my (3) GPU's utilization running Mandelbub's default screen, at it's default res, is 64% - 70% - 59% utilization.


I did again reboot fresh, and only ran the Performance Monitor, and Mandelbub...
Let the default screen come up, and used the advanced 'f' key to go 1920x1200 res. (On my system.)

I guess like games, res is important, and let me see more cores would be used in the system deemed it necessary.
My CPU utilization now looks like this with Mandelbub's default screen at 1920x1200. (Getting 37.96FPS.)

For trivia my (3) GPU's utilization running Mandelbub's default screen, at 1920x1200 res, is 72% - 82% - 67% utilization.

Ya know freaky, it looks to me like your CPU does a better job splitting it's work up amongst it's cores, than my Q6600 does...

Try out the '3' too, for 3D Mode. Good if you want a monster load on the system quick... It cut's your frame rates in 1/2, probably as expected.

I also enjoy taping the space bar, for animation mode.

Sure...

Q6600@3.81GHz, (1) 295@ 621/1512/1152.

I pointed my PhysX back to the 295 in Nvidia Control Panel...

Fired up (3) instances of GPU-z to verify processing activity...

My 280 was indeed out of the equation, as a processing factor...
Still getting 74.40 FPS at the default screen, default res.




Next I shut down all (3) instances of GPU-z, then used the 'f' key to go to 1920x1200 res.
I still get 28FPS.


So for me, default view, at default res: (1) 280 adds 19FPS.    (Was 93, now 74, diff=19.)

And the default view, at 1920x1200 res: (1) 280 adds 9FPS.   (Was 37, now 28, diff=9.)

When I installed my new WHQL driver, I kept some selections off...

That could be part of it?



I know Ambient Occlusion has a performance hit on the system...

I also am still impressed with what the 280 added...

Not bad at all...

Does this count?



Hehe...sorry...I've been following your threads Talon (good stuff btw).  I'm encoding a video for the Linux forums (all 10 of us down there) and I couldn't resist (yes...there are much better sys monitor tools available for Unix I just am...well...err..old).

Hey talonman,
Your right with furmark only running 1cpu core.
But i can't get both my gpu's working in sli with furmark only one. We both have the same version of furmark.
And i have the sli profile from nvidia not from evga. Did you force a sli profile for furmark?
I found the problem talonman, my one gpu was very unstable with my oc. I guess causing so much instability it shuts the one gpu down but not the other gpu. Its funny how i could run mandelbulb with both gpu's and have a stable OC but with furmark its a different story. This must be the problem i have with downclocking in some games. I guess that's the difference with different applications. I'm not sure yet but so far that's what its pointing to.
I'll let you know how it turns out.

ok i have a question for you, furmark is only running one thread with my cpu. Would that be considered half a core? Or wouldn't two threads be a full core? Wouldn't the cpu become my bottleneck? You would be using a full core and i'm only getting half a core out of furmark. I'm guessing that's why the gpu is running at full load with this application. And applications using quad cores doesn't need the gpu to run at full load. That would be common sense lol. But applications like furmark would be a major cpu bottleneck for 8 threaded cpu's.
Furmark starting

freakysqeeky

ok i have a question for you, furmark is only running one thread with my cpu. Would that be considered half a core? Or wouldn't two threads be a full core? Wouldn't the cpu become my bottleneck?
Furmark starting  

1/2 a core... I guess?

We do know 1 thread for sure, and if Furmark can pin your processor, my Q6600 doesn't have a prayer of getting off 100% utilization.

I think it just takes your 1 core to 100%, no matter how fast it is?

Time for a look at FluidMark version 1.1.1 (The PhysX Benchmark.) 

I selected 16 x MSAA in this test...

System: (1) Q6600 @3.81GHz, (1) 295 in SLI Mode, and (1) 280 operating in dedicated PhysX mode:


Sad that only 1/2 of my 295 gets used, and that my 280 is only at 19% utilization...

So both FluidMark, and 'Crash-n-Burn', both PhysX benchmark apps mind you, barely use the PhysX processor at all... 

I almost stand in disbelief now, that this small GPU utilization is the total sum of 'Nvidia's huge advantage', in Vantage with PhysX enabled:

You would think a CPU could deal with that...

I wish we had PhyX Benchmark apps that demanded more utilization from your PhysX GPU!
Games do...

For trivia, keeping the same res as the above post, 1280x1024:
When 4 X MSAA was selected, 1/2 of my 295's utilization was 54%, and 280's was 19%.
When 8 X MSAA was selected, 1/2 of my 295's utilization was 83%, and 280's was 19%.
When 16X MSAA was selected, 1/2 of my 295's utilization was 94%, and 280's was 19%.

Looks like I had plenty of CPU for this test...

Good info...

Thanks for the post.

BatMan Arkham Asylum v1.1

In game settings:


System: (1) Q6600 @3.81GHz, (1) 295 in SLI Mode, and (1) 280 operating in dedicated PhysX mode:


Decent use on both sides of the 295, but not has high of a utilization as Cryostasis...

The 280 is being utilized in this game for PhysX pretty consistently as you play, and as far as utilization goes, I guess it's about as high as they come? 

Outstanding job of workload distribution on all 4 cores of my CPU. (Now that's sexy!) 

Running 3 instances of Folding@home  work unit #5781

EDIT... Heres a screeny with the correct Gpu-z

Doh ,i have it ,i just selected the wrong shortcut.

I'll edit my post