Boinc only using 70-80% of gpu

[phillrulz]

I have a 780ti and want to realy push through these WU however Boinc is only using 70-80% of my gpu and processor isnt being touched as im using a gpu based WU (GPU grid) any reason why its not using all of my gpu ?

[Sauron]

Probably it's due to the fact that the computing cores on a 780ti are limited, hence not the whole card is being used. I wouldn't worry about it, it's probably using all it can use already.

[phillrulz]

Probably it's due to the fact that the computing cores on a 780ti are limited, hence not the whole card is being used. I wouldn't worry about it, it's probably using all it can use already.

thanks just i know with F at home it would max out

[Whaler_99]

Also some "work" will stress the card more then others. Also see this a bit with FAH.

[phillrulz]

Also some "work" will stress the card more then others. Also see this a bit with FAH.

yea recent boinc WUs have been messing with the estimated time, atm i have a short 2-3 hour it says estimated time left 3:26 however its more likely to complete in about 50 mins

[alpenwasser]

This is quite normal.

When I started to run GPUGrid on my Titan, GPU utilization was

somewhere between 40% and 60%.

Now, there is something you can do, but it comes with two significant

caveats.

Usually GPUGrid will only run one application at the same time per

GPU, but you can configure it in such a way as to increase that number.

For that, do the following: Go to your BOINC data directory. See

this link for where that is.

Once you're there, go into the directory "projects\www.gpugrid.net".

Then create a file (by default it doesn't exist I believe) called

app_config.xml. Open with a text editor.

This is the content of my app_config.xml for GPUGrid:

<app_config><app><name>acemdshort</name><max_concurrent>10</max_concurrent><gpu_versions><gpu_usage>.4995</gpu_usage><cpu_usage>1</cpu_usage></gpu_versions></app><app><name>acemdlong</name><max_concurrent>10</max_concurrent><gpu_versions><gpu_usage>.4995</gpu_usage><cpu_usage>1</cpu_usage></gpu_versions></app></app_config>

Now, I must admit it's been a while since I've set this up, so

I don't remember all the specific reasons for why I've chosen

the numbers I have here (and you could change some things without

making a difference, e.g. the max_concurrent thing).

However, if I recall correctly, it works something like this:

The max_concurrent indicates the maximum number of that application

which are allowed to run at the same time, naturally. The

gpu_usage number does the following:

(number of GPUs) divided by GPU_usage gives you the number of

applications which will run at the same time on your machine.

So, in my case:

1/0.4995 ~= 2.002 ==> 2 applications will run at the same time,

provided the max_concurrent flag is not below this number.

If I had two GPUs, I'd be running four GPUGrid applications

at the same time (two per GPU), and so on.

Using this, you should be able to run two PGUGrid applications

on your GPU at the same time. I have currently running two

apps concurrently, and my GPU utilization hovers between 70%

and 90%, usually 80% (as Whaler has said: not all projects are

the same in this regard).

However:

Caveat #1: You may lose GPUGrid bonus credits. If I recall correctly,

GPUGrid expects its results back within 24 hours to award full

bonus credits, and those credits are rather significant (at least it

was like this the last time I checked a few months ago). And if

you increase the number of applications running concurrently on a

GPU, the time needed to finish them will go up significantly.

Currently, a large GPUGrid unit takes between 9 and 16 hours on my

machine, depending on the unit. If your GPU is already utilized at

80% or so, it might be that you'll go above the 24 hour threshold

if you run another one concurrently. You'll need to test this out.

If you do go above that, I'd say it's not really worth doing TBH.

They give those bonus credits for a reason (i.e they want their

results back fast), after all.

Second caveat: Law of diminishing returns. For testing purposes,

I tried out what would happen when I ran three units in parallel

on a single GPU, and GPU utilization only went up to ~90%, while

calculation times went up significantly. So, going from one to

two units gave me about an additional 30% GPU utilization, while

going from two to three only yielded another 10% or so (again,

this will depend on the specific unit being worked on).

Bottom line: Running a config which reaches 100% GPU utilization

is not reasonable, but you might be able to run two units in

parallel.

Side note: Be very careful with typos in the app_config.xml file,

specifically the opening and closing tags. XML is very strict

about this.

[phillrulz]

-snip-

 

 

Great read i think ill stick to the 80% tho thanks for the indepth view it nice to see a balanced argument.