Brp7/MeerKat "not enough gpu memory"
12 Oct 2023 19:08:35 UTC
Topic 230190
(moderation:
I "finally" got this system to recognize the Internal gpu under Boinc (repository version).
When I try to run a Brp7/MeerKat tasks it says "task paused, not enough GPU memory".
I have played with the "UMA Buffer" which seems to be the only Bios level way to try to allocate more GPU Memory.
No luck. I even tried running the task with no other cpu task present. No luck.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
Tom M wrote: I "finally" got
)
...and your question is...???
...and you have "played with the "UMA Buffer" ", meaning what...??? What did you do to play with it?
You have several error's in your "Stderr output" file, and one with a code
Proud member of the Old Farts Association
The out-of-memory message is
)
The out-of-memory message is misleading - the real error is "
Couldn't create OpenCL command queue (error: -6)!"
And this has everything to do with a parameter in boinc-client.service called ProtectSystem. This parameter was changed/added circa boinc release 7.18 and it needs to be turned off/disabled. There is a thread with more info about this issue and near the end of the thread, the solution.
Also boinc 7.18 isn't intended for Linux and should not be in the repo - it was an Android release. Please use the 7.20.5 from the Costamagna ppa.
this is you actual error from
)
this is you actual error from the only BRP7 task you ran:
the verbiage about not having enough memory I'm certain is a byproduct of this earlier failure. it could be a driver issue, or it could be that the application just isn't coded in a way that the APU supports.
I would recommend not using the APU for this.
_________________________________________________________________________
---repost from other
)
---repost from other thread---
Been running only Meerkat x5 units for a few days now.
Once in a while there's a unit running running lonely with around 800s-900s being a bit off, they were validated.
Didn't see any of those on the screen myself.
Today I had a compute error without me tinkering:
https://einsteinathome.org/workunit/766701870
I'm running BRP7 Meerkat test app.
Not running petris app.
Ends with error 700 (and 1008), seems to be cuda illegal memory access?
Don't know what really is happening there, lets post it in a Meerkat thread.
---/repost from other thread
I didn't run into a VRAM shortage, but error 700 during CUDA device seems to be an illegal memory access.
Maybe memory errors are related, I do run the Meerkat BRP7 test app atm.
since it's only one i wouldnt
)
since it's only one i wouldnt worry about it.
maybe caused by a random memory error? any overclocks? maybe back it off a tad if so.
but again, unless it becomes a persistent thing, i wouldnt worry about it.
_________________________________________________________________________
There's wasn't too much of a
)
There's wasn't too much of a clock on it, but there is upthrottling during the day.
This card throttles in steps of 15MHz, visible in the NVsettings gui.
After I set the clock it can go up and down a bit + or - 15MHz or usually 0 throttle.
Been watching it regularly yesterday and it went up 30Mhz and rarely even up 45MHz, think it was temperature throttle. There has been an increase in invalids to around 4%(up to 6%?) recently (with 60-70% validation checks). Clocked it down with 45 now.
Then I changed fanspeeds and see the card throttle down when either warm and cold, The highest throttle is around 55C temperature. While I was at it, why not put the fans on 70 and wait, 80, 90 and 100% to check for noise and temps.
Hitting that 100% awaiting the spin up, the fans seemingly overspun and the message in NVsettings became Unsupported fanspeed, compute errors in boinc, the fans kept maxing out not responding to try and set it lower.
My compute errors on 16 Nov 2023 10:07:55 UTC were self induced.
Rebooting worked.
HWpecker wrote: ---repost
)
One person said the new tasks take 4.5gb of ram and won't run on 4gb gpu's any more.
mikey wrote: One person said
)
wrong app and outdated info.
HW is talking about the BRP7 app, which uses a small amount of VRAM (~1GB)
the app that used "4.5GB" was the O3AS app, and that was the old version of the app. Bernd released a new application and tasks that use about 2-2.5GB now.
_________________________________________________________________________
Ian&Steve C. wrote: mikey
)
Thank you very much, I knew Bernd had fixed the new ones but forgot about the rest.