Forum > GPU crunching

Modified SETI MB CUDA + opt AP package for full GPU utilization

<< < (40/58) > >>

Leopoldo:

--- Quote from: Maik on 15 Jan 2009, 04:10:23 am ---give my script a chance ;)
still have freezing tasks at different runtimes ... without the script i would have same problems lik you ^^
Raistmer's v5b + script = well crunching on a 9600GT

--- End quote ---

Agreed. Thanks both!

Richard Haselgrove:

--- Quote from: Jason G on 15 Jan 2009, 05:06:35 am ---Cheers both.   I will add script first, then try underclock, then overclock.  My temps are higher because of climate here though test stable with ATI tool and RthDrible.  I am most prone to think low speed is the culprit (Nvidia choosing too big a granularity for the parallelistion), as in the past I have my 8600GT's run @100C no problems under stress (68 degrees fully loaded now).

[If underclock fixes it, then I blame temps, If Overclock Fixes it then I blame Nvidia's choice of target cards]


--- End quote ---

I'm quite convinced it's a programming bug in the SETI application.

Have a look at my bug report from last night.

Normally, my 9800GT processes a mid-range AR in 22-23 minutes, regular as clockwork. It's still running, still doing that this morning. This is with the Berkeley Beta v6.07, on official download - the card is rated as fast enough by NVidia to be entrusted with the CUDA app.

At VLAR - different story. Took 20 minutes to reach 4%, an hour to reach 35%. Then just idled (endless loop?) for three hours, didn't even checkpoint. I suspended the task, and re-started the BOINC daemon to get it to recognise the suspension - normal work restarted. No reboot, no driver crash.

Everything is in that upload - datapak, checkpoint, wisdom, client_state. See if you can find the bug before Eric K does.

Jason G:
Nah, steering away from looking at cuda code for the time being, partly because I want to skip it and go straight to OpenCL when vendor SDKs are available.  It has plenty of heads looking at it and seems to be getting better gradually.  I am sure there are multiple issues at play, and one of them definitely seems to be speed. cranking up the 9600GSO   (along with the fan), and checking for artefacts in rthdribl and ati tool has me up to 68GFlops estimated using that formula, and things indeed seem a little better than stock 48 GFlops estimate.  Will see what I can get stably without producing too much heat or generating artefacts, and comment on what I find.

Jason G:
Found card stability test that works for me, Will run 1 hour ati tool artefact scan.

card is 9600GSO,
Have settled for 56Gflop conservative OC ( using estimate formula of ClockRate X MultiProcessors X 2857 )
Clocks Obtained with ATi-Tool 'Find Max Mem' and 'Find max Core' separately, after reducing safety limits to keep in range.  No atrefacts yet, card nice and warm, 500frames per second in RthDrible Large, no pausing. (Then installed in RivaTuner afterward)

Raistmer:
Ok, here modded build with VLAR bug-fix incorporated.
It should do correctly all AR ranges (please,  post erros if they will be).
But be prepared to VLRT (very long run time) for VLAR tasks.
I'm not sure that it's worth to do VLAR on GPU at all with current algorithm. So, you always can reverse to V5b and leave VLARs to hosts with CPU-based opt apps.
But before doing this compare wall clock times for your own hosts (!) I use pretty low-end GPU, your experience can (and will ;) ) be different.

ADDON: I didn't test this build much so if you experience some new errors - just revert back to V5b or switch to 6.07 and forthcoming 6.08 on beta site.


[attachment deleted by admin]

Navigation

[0] Message Index

[#] Next page

[*] Previous page