GPU load monitoring tool Now available!

Answer from the programmer of GPU-z, w1zzard:

“I’m using the official interface provided by nvidia in their nvapi.”

NOTE: SiSoftware Technical Support does not allow me to upload pictures to them, so I will be using this thread as a picture dump, to help better explain my setup…

(Sorry if your sick of my pictures) :">

Answer from Sandra Support:

"Date: 2009-12-07 10:25:28

Name: SiSoftware Technical Support

Message:

The processing is done on the card(s) that you select from the 2nd combo-box (the 1st one selects test type). So whatever the interface returns that is what is used.

You can thus select to use both cards simultaneously or each individually. If the OpenCL driver supports CPU, you can also select CPU+GPU or CPU / GPU individually.

AFAIK the dedicated PhysX card is used for PhysX when the game/app requires PhysX processing; it is available for other work at other times.

For all interfaces you may need to connect/enable an output on the other cards to make them available for processing (e.g. connect a monitor, VGA dongle and extend desktop).

Best regards,

The SiSoftware Technical Support Team"

PS. The latest version of Sandra is version 2010 (16.xx)."

My response to Sandra Support:

But in the pictures shown in the the various threads, show that I only have my 295 selected, when processing is getting done on the 280. That is unexpected!

If you don’t select the 280, processing shouldn’t happen on the 280, yet it does.

For trivia, I do have a display connected to every card.

(I am a Folder, and I already had to do that to be able to load up 3 instances of folding.)

I have yet to see how to select multiple GPU’s, to run at the same time in (free) Sandra Light.

Their response:

Both cards should be visible as well as the “combined” card that represents both.

Sandra Lite has exactly the same benchmarks as all the other types, the only difference is lack of other modules.

Let’s try:

  • uninstall the drivers

  • install latest 195.62 drivers

  • make sure both adapters are enabled in Windows with one display each

  • make sure SLI is disabled

If nothing works I will ask our nVidia contact to have a look.

While we have not tested your specific configuration, we have tested:

  • 2x GTX 260

  • GT 9600 + GT 9500

  • on board 9400 + 9500 dedicated just fine.

Thus there is no reason for things not to work on your system.

Best regards,

The SiSoftware Technical Support Team

PS. The latest version of Sandra is version 2010 (16.xx).

Both cards should be visible as well as the “combined” card that represents both.

These are my available selections for OpenCL: (The OpenCL test has an extra 295 option in there… 3?)

These are my available selections with CUDA: (The CUDA test has an extra 295 option too.)

These are my available selections with Compute Shader: (The Compute Shader test just has (2) 295’s to select from.)

I see no option for the “combined” card that represents both in any of the tests.

(I might be missing it too?) :)

My device manager looks like this, and am running with SLI and PhysX enabled:

- install latest 195.62 drivers

I am currently running on them, and working well for me.

Note that it is in both the CUDA and Compute Shader test’s, that my 280 receives the most processing action.

[b]- make sure both adapters are enabled in Windows with one display each

  • make sure SLI is disabled[/b]

I hate to run with SLI disabled, I believe the CUDA test is using it to help my score. I will however try that, to see it my dedicated PhysX processor will no longer receive any processing action, when I have my 295 selected as the GPU to be tested.

Note: That I do have valid display connected to all 3 GPU’s

One side of my 295 is connected to my Gateway main FPD.

The other side of my 295 is connected via HDMI, to a small LG TV. (works fine)

And my 280 is connected to a second input on my display, with an adapter.

The 280 works well to if I activate it in Nvidia Control Panel, and select VGA as input on my display.

I was running set for Single display, Performance mode in the Control Panel.

This is the best way to show my load distribution.

I fired up 3 instances of GPU-z, for a better graphical representation of what my 3 GPU’s processing activity was, while each test was running.

NOTE: Every test was run with the default 295 selected, when Sandra initially starts. I only changed the test, not the GPU.

OpenCL Test:

CUDA Test:

[b]

Compute Shader Test:[/b]

When support looks at this, note that my event viewer doesn’t like Sandra too much!

(I have no idea what that means.) External Media

I still find it worth noting, that when I ran ATI Tool with my current config, having both SLI and PhysX enabled…

Processing load was kept to only my 295, and no utilization on my 280 at all.

One 1/2 of my 295 reads 31%.

The other 1/2 of my 295 reads 94%.

And my 280 operating in Dedicated PhysX Mode, reads 0% load, just as it should.

The Final Outcome

Date: 2009-12-07 23:32:08

Name: SiSoftware Technical Support

OK, based on the pictures:

  1. Your 1st card is the 295(both GPUs on it) + 280

  2. Your 2nd card is the 295 - 1 GPU

  3. Your 3rd card is the 295 - 2 GPU

  4. Your 4th card is the 280

Thus if you select (1) you are running on ALL 3 cards, both parts of the 295 + 280.

If you select 2 or 3 you are running 1/2 of the 295.

If you select 4 you are running on the 280.

The issue here is that you have a 295 which through CUDA / OpenCL are ennumerated as indivudual GPUs, thus 2 entries for each GPU on it.

DirectX/Compute Shader uses SLI thus shows 1 card (it combines the 2 GPUs on the 295 as you know).

Sandra knows the 295 is X2, thus it does not say 2x GTX 295 but shows you 2x shaders/memory/etc.

This is purely cosmetic - if you select (1) we run on all cards in parallel, with work distributed depending on their power. I will see what we can do to fix it.

The work-allocator works out how much to give to each card, e.g. if one is 2x as fast it gives 66% of the work and 33% to the other.

I will ask for lab time if possible, but there are not guarantees that we will get it or this particular configuration would be available for testing.

Please note that as you are using the free Lite version and technical support is also free, while we do investigate all bugs, we may not pursue all to a resolution if we cannot work out what’s wrong and how to fix it.

Best regards,

The SiSoftware Technical Support Team

My response:

Outstanding!

Thank you for your time!

I would love if you did get some lab time for us Nvidia guys…

I would like to see a much better utilization of our GPU’s.

Between the intermittent processing issues causing GPU-z to not report correctly, and some apparent low utilization when it does report correctly, I cant help be feel our scores could be better.

As you have pointed out to me, that the light and full blown Sandra run the exact same test, both apps would benefit from further optimizations.

They may like to see my data in the provided thread.

Thanks again.

Name: SiSoftware Technical Support

As I mentioned, we hope to be working with nVidia as we work with other manufacturers to ensure we are extracting the best performance of the devices.

If we do get any suggestions/fixes, we will release a Beta version for people to test.

BTW, make sure you are using version 16.11 with support for CUDA 3.0.1 and OpenCL drivers 162.65; the original 16.10 used CUDA 2.3 and previous drivers/OpenCL.

Best regards,

The SiSoftware Technical Support Team

PS. The latest version of Sandra is version 2010 (16.xx).

Thanks again, I will do just that… :)

Update: I am on Version 2010.1.16.10 of Sandra Light.

FYI - The latest version 2010.1.16.11 of Sandra Light is here…

http://www.sisoftware.co.uk/index.html?dir…n=sware_dl_3264

I will be sure to update my version.

After running some tests on version 2010.1.16.11, I would have to say all (3) of my GPU’s utilization looks the same as the pictures posted above.

(Menu’s are the same too.)