Advanced search

Message boards : Graphics cards (GPUs) : GeForce 780 utilization only 85-90%, even with 2 projects at the same time

Author Message
GrayFox
Send message
Joined: 26 Mar 17
Posts: 4
Credit: 2,942,450
RAC: 8,324
Level
Ala
Scientific publications
wat
Message 46792 - Posted: 29 Mar 2017 | 10:42:09 UTC

Hello,
i had low gpu usage.
I tried add to app_config - max_concurrent 3 and it worked, it increased usage from 74% to 85-90% and now 2 projects are crunched at the same time, instead of 1, but still not 3.
Gpu usage 85-90% is still not effective, while crunching einstein project, my gpu is at 99% constantly.

Next thing i tried: run two clients using program BoincTasks and i ran to several problems:
-BoincTasks doesn't load app_config, so only one task ran per client, that means i would need 3 clients and than i hopefully could utilize my gpu.
-Tasks got aborted all the sudden, no idea why.
-I don't know if i even ran that second client correctly.

There are almost no guides and i was only get partial information from the forums, which context i don't understand.
How do i get my gpu usage to 99% with this project please ?

PappaLitto
Send message
Joined: 21 Mar 16
Posts: 263
Credit: 1,203,497,131
RAC: 4,734,330
Level
Met
Scientific publications
watwat
Message 46793 - Posted: 29 Mar 2017 | 12:03:46 UTC

Some of these GPUGrid Work Units require a fast CPU to feed the GPU quick enough to enable high GPU utilization. A CPU with high IPC (Instructions per clock) and high clock speed are desirable. Not all GPUGrid Work Units require a fast CPU as the more atoms simulated the longer it takes the GPU to produce a frame and the less dependent it becomes on the CPU. The problem is the smaller atom simulations that have high draw call. For quickest return rate, which is more important to the researchers than utilization, I would run 1 Work Unit per GPU even if the utilization is less than optimal. You can improve your GPU utilization by overclocking your CPU, switching your crunching PC to a non-WDDM Operating system such as Windows XP or Linux.

GrayFox
Send message
Joined: 26 Mar 17
Posts: 4
Credit: 2,942,450
RAC: 8,324
Level
Ala
Scientific publications
wat
Message 46795 - Posted: 29 Mar 2017 | 13:42:44 UTC - in response to Message 46793.

Okay good to know, thank you.

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 1837
Credit: 10,434,859,894
RAC: 8,761,703
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 46796 - Posted: 30 Mar 2017 | 0:13:30 UTC - in response to Message 46792.
Last modified: 30 Mar 2017 | 0:14:29 UTC

Hello,
i had low gpu usage.
I tried add to app_config - max_concurrent 3 and it worked, it increased usage from 74% to 85-90% and now 2 projects are crunched at the same time, instead of 1, but still not 3.
GPUGrid limits the tasks for a host at 2 workunits per GPU, so you can't have 3 per GPU.

Gpu usage 85-90% is still not effective, while crunching einstein project, my gpu is at 99% constantly.
GPU usage shown by 3rd party tools are not the CUDA core usage. You can have a more precise CUDA core usage reading by measuring the power consumption of your PC (or the temperature of your GPU), it will be higher while you crunch GPUGrid on 74-80% GPU usage then crunching Einstein@home on 99% GPU usage (this project has an older CUDA client).

There are almost no guides and i was only get partial information from the forums, which context i don't understand.
How do i get my gpu usage to 99% with this project please ?
See this post for tips on maximizing GPU usage.
BTW it would be easier to give you tips if you would unhide your computers on GPUGrid, or post the details of your configuration here.

GrayFox
Send message
Joined: 26 Mar 17
Posts: 4
Credit: 2,942,450
RAC: 8,324
Level
Ala
Scientific publications
wat
Message 46798 - Posted: 31 Mar 2017 | 9:27:34 UTC - in response to Message 46796.
Last modified: 31 Mar 2017 | 9:41:09 UTC

Posted by PappaLitto
I would run 1 Work Unit per GPU even if the utilization is less than optimal. You can improve your GPU utilization by overclocking your CPU, switching your crunching PC to a non-WDDM Operating system such as Windows XP or Linux.


After i changed cpu utilization to 12,5%, or it should be 13.5% ?
I got 90% GPU usage, even while i am crunching only 1 task at the same time.
So it doesn't seem, you get any advantage from crunching 2 tasks at the same time, especially if GPU usage doesn't matter anyway.
But TDP does, which is the same - 60% max, either i crunch 1, or 2 tasks at the same time, i don't know if there is something else to be done about that here.

You recommend Linux, but Retvari Zoltan says in this topic, that he has only 7-8% cpu usage in Ubuntu 16.04, due swan_sync is not supported there
Ofc there are many distros, i will try some probably.

But gpu utilization is gived by application programming, or it can be something on Nvidia's side i guess and 780 isn't most up to date card.
Probably not much i can do here.

EDIT: God :O, in afterburner, there is a feature - temp limit, i am used from gigabyte oc guru to target temp, which is only preference not a limit, well tdp went up to 6%, will tweak that further.
Ye it increased perfomance, well...

PappaLitto
Send message
Joined: 21 Mar 16
Posts: 263
Credit: 1,203,497,131
RAC: 4,734,330
Level
Met
Scientific publications
watwat
Message 46800 - Posted: 31 Mar 2017 | 11:40:22 UTC - in response to Message 46798.

Posted by PappaLitto
I would run 1 Work Unit per GPU even if the utilization is less than optimal. You can improve your GPU utilization by overclocking your CPU, switching your crunching PC to a non-WDDM Operating system such as Windows XP or Linux.


After i changed cpu utilization to 12,5%, or it should be 13.5% ?
I got 90% GPU usage, even while i am crunching only 1 task at the same time.
So it doesn't seem, you get any advantage from crunching 2 tasks at the same time, especially if GPU usage doesn't matter anyway.
But TDP does, which is the same - 60% max, either i crunch 1, or 2 tasks at the same time, i don't know if there is something else to be done about that here.

You recommend Linux, but Retvari Zoltan says in this topic, that he has only 7-8% cpu usage in Ubuntu 16.04, due swan_sync is not supported there
Ofc there are many distros, i will try some probably.

But gpu utilization is gived by application programming, or it can be something on Nvidia's side i guess and 780 isn't most up to date card.
Probably not much i can do here.

EDIT: God :O, in afterburner, there is a feature - temp limit, i am used from gigabyte oc guru to target temp, which is only preference not a limit, well tdp went up to 6%, will tweak that further.
Ye it increased perfomance, well...


From my experience, and Zoltan can chime in, I've noticed that "GPU Usage" is the primary contributor to quicker turnaround of the WU. Power consumption/TDP seems like it would matter but from my experience different GPU manufacturers scale their TDP % differently. For example, my 980ti uses 85% max TDP at 94% usage while my 1070s which are AMP Extreme cards meant for overclocking are at no more than 65% TDP at 93% GPU usage. My 690s are thermally restricted because of the smaller heatsync on each GPU, therefore the TDP to run max clock is actually 110%+ at 95% GPU usage. In my opinion you should focus on ways to improve GPU utilization through CPU and software optimizations. You're getting 90% utilization with that one particular WU but the next it could be 70%. You want to make sure your CPU is at max boost clock at all times to ensure the information returns to the GPU as fast as possible. I haven't tried Linux myself as I didn't have time to get the GPU driver up and running so I decided to go with Windows XP myself.

GrayFox
Send message
Joined: 26 Mar 17
Posts: 4
Credit: 2,942,450
RAC: 8,324
Level
Ala
Scientific publications
wat
Message 46802 - Posted: 31 Mar 2017 | 13:28:44 UTC - in response to Message 46800.

Apparently there is gpu usage preference for this project, it calls "Maximum CPU% for graphics", i would never expect it would be named like that.
I thought, there is no option for controlling gpu usage with this project.
My bad.

I can't OC, because i have cheapest mobo.

Ye i know about unparking cpu cores and disabling power saving features in bios.

My cpu is currenttly constantly running at 3.67 ghz, i can get up to 3.9 ghz with turbo and either eist,c1,c6, or thermal function cause it run at 4.12 ghz tops, which is ackward, but it adds latency and cause inconsistent clocks, it may be worse than run cpu at 3.67 ghz constantly, i don't know.

I will overclock my gpu yet gradually, it is all good now.
Thanks for help btw.

Post to thread

Message boards : Graphics cards (GPUs) : GeForce 780 utilization only 85-90%, even with 2 projects at the same time