Weird issue with 7970

nanoprobe
nanoprobe
Joined: 3 Mar 12
Posts: 40
Credit: 12540756
RAC: 0
Topic 225074

I recently put an old 7970 GHz. edition back into service. I wanted to test run it here to see if was still in working order. I have it set to run 2 tasks simultaneously. The tasks start, run for about 18 minutes, get to about 80% completion and then the weirdness begins. The progress bar goes back to about 3.3% complete and basically doesn't progress any more from there. I let 1 run for an additional hour but it only progressed to 12%. This is a Win 7 64 bit OS with the latest driver. My 270x on the same OS and my GTX 1060 on Linux run the tasks without issues. The 7970 ran some GPU beta tasks on WCG without issues. Anyone have any idea as to what's going on?

archae86
archae86
Joined: 6 Dec 05
Posts: 3157
Credit: 7220594931
RAC: 938944

The easiest answer would be

The easiest answer would be that it is a slow card, which you are further slowing by asking it to run 2X.  That the speedy initial progress is an illusion we sometimes call "synthetic progress" displayed before the application actually reports a useful progress point to BOINC.  That you have consistently terminated things long before it could finish.

Running 2X on Einstein Gravity Wave on a card reporting as 3072 MB is highly unwise.

So unless something else turns up, I suggest:

1. run 1X (initially you could do this just by suspending all but one task.
2. be patient--let it try to finish, for at least a half day.

Maybe there is in fact a "weird issue", but I suggest checking out the simple possibilities first.

GW tasks have considerable variation in GPU RAM requirement.  Getting some to run at a given multiplicity is no guarantee that the full range will run.

mountkidd
mountkidd
Joined: 14 Jun 12
Posts: 176
Credit: 12566162555
RAC: 8035578

See this post  from Gary.  

See this post  from Gary.   GCN1 cards can't run GW.  However, GR is ok.

nanoprobe
nanoprobe
Joined: 3 Mar 12
Posts: 40
Credit: 12540756
RAC: 0

I must be missing something.

I must be missing something. The 270x only has 2 GB of memory and it runs them 2 at a time. Shouldn't the 7970 with 3 GB be able to run them? Aren't they based on the same architecture? Tried running only1 on the 7970 and it still had the same issue. Might be some other issue with that card.

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5872
Credit: 117545043438
RAC: 35300733

nanoprobe wrote:The 270x only

nanoprobe wrote:
The 270x only has 2 GB of memory and it runs them 2 at a time.

BOINC lists it as an R9 200 series.  I looked at your tasks list for that machine and, on the very first page, found a completed task, along with a whole bunch of aborted tasks.  By clicking on the TaskID link for it, you get to see the info that was returned to the project.  Here is a particular line from that log info:-

OpenCL Device used for Search/Recalc and/or semi coherent step: 'Tonga ... (2048 MiB)'

Tonga is GCN 3rd gen which is why that card could work.  A long time ago I tried a HD7950 (Tahiti) so I know for sure that unless something has changed with the app more recently, there's no chance that GCN 1st gen cards will work.  There have been many examples since then (the very latest is yours :-) ) that continue to show the same behaviour.  There's never any real progress (just fake progress) and eventually a maximum time limit is exceeded and the task is terminated (if you wait that long).

Cheers,
Gary.

Sid
Sid
Joined: 17 Oct 10
Posts: 164
Credit: 970063529
RAC: 419195

Gary Roberts wrote: Tonga is

Gary Roberts wrote:

Tonga is GCN 3rd gen which is why that card could work.  A long time ago I tried a HD7950 (Tahiti) so I know for sure that unless something has changed with the app more recently, there's no chance that GCN 1st gen cards will work.  There have been many examples since then (the very latest is yours :-) ) that continue to show the same behaviour.  There's never any real progress (just fake progress) and eventually a maximum time limit is exceeded and the task is terminated (if you wait that long).

The most interesting thing about 7970/280x and GW tasks -  MSI 280x with 6Gb has the same issue - "There's never any real progress (just fake progress) and eventually a maximum time limit is exceeded and the task is terminated (if you wait that long)."

Tried several times in 1X mode.

Ian&Steve C.
Ian&Steve C.
Joined: 19 Jan 20
Posts: 3945
Credit: 46749322642
RAC: 64134435

maybe nanoprobe is mixed up

maybe nanoprobe is mixed up about what his card actually is?

the only R9 200 series card that's Tonga looks to be the R9 285, not the 270x (which was Pitcairn).

_________________________________________________________________________

nanoprobe
nanoprobe
Joined: 3 Mar 12
Posts: 40
Credit: 12540756
RAC: 0

7970 seems to be running the

7970 seems to be running the GR tasks OK. 2 complete in about 18 minutes. Just 1 more question. I have the GR tasks also running on the 270x. I have the website settings set @ 2 tasks simultaneously. Can the 7970 run 3 or 4 at a time and if so can I override the website settings for the 7970 with an app config file in the project folder? Don't want to run more than 2 on the 270x. Thanks

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5872
Credit: 117545043438
RAC: 35300733

nanoprobe wrote:... Can the

nanoprobe wrote:
... Can the 7970 run 3 or 4 at a time ...

Did you read the comment that Mountkidd linked in his earlier response?  That specific question was already answered there.

I should point out that if you go above 2x multiplicity, you should carry out your own experiments first to really be sure if the outcome is the same for your particular system.

nanoprobe wrote:
... can I override the website settings for the 7970 with an app config file in the project folder?

Yes.

Website preferences for multiple concurrent tasks are overridden by the contents of a local app_config.xml.  I use such a file in every host of mine.  I can configure multiplicity on a per host basis and have the change acted on immediately using the BOINC Manager's 'reread config files' option.  With website settings, the change is only transmitted to the client on receipt of new work (not just a simple update).

Cheers,
Gary.

mountkidd
mountkidd
Joined: 14 Jun 12
Posts: 176
Credit: 12566162555
RAC: 8035578

Along the way, I ran 2x/3x

Along the way, I ran 2x/3x with my 7970.  3x did provide a very modest improvement over 2x, but the task times had a high degree of variability.  Changing to 2x yielded task times that were much more consistent.  My interpretation was that at 3x there was competition for GPU resources resulting in interference.  My 7970 is OC'd.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.