Seti@Home optimized science apps and information
Optimized Seti@Home apps => Windows => GPU crunching => Topic started by: SilverL on 17 May 2009, 12:54:55 pm
-
I just upgraded to the latest drivers and for some reason now my cuda tasks wont switch over from the CPU. I have the same setup with a 9600M GT and it runs fine. Can anyone help?
-
I just upgraded to the latest drivers and for some reason now my cuda tasks wont switch over from the CPU. I have the same setup with a 9600M GT and it runs fine. Can anyone help?
At least link to your host required, don't you think so? ;)
-
Here you go.
http://setiathome.berkeley.edu/show_host_detail.php?hostid=4819179
It used to switch to GPU before, I dont remember the driver though.
-
Total GPU memory 268435456 free GPU memory 165355520
Restarted at 3.05 percent.
Cuda error 'cudaMalloc((void**) &dev_outputposition' in file 'd:/BTR/SETI6/SETI_MB_CUDA/client/cuda/cudaAcceleration.cu' in line 369 : out of memory.
Red value seems not enough for app initialization.
Try to reboot and see what value will be printed in this field.
Is it possible to increase available memory for this GPU? (I suppose it's embedded one? )
-
Not even after a restart did they go to the GPU. So I guess theres no chance of using that card anymore for crunching. Its all ok though. Thanks for the help.
-
Not even after a restart did they go to the GPU. So I guess theres no chance of using that card anymore for crunching. Its all ok though. Thanks for the help.
If you want to debug this problem little more try to reduce screen resolution to minimal value and color bittness. And see how reported Free GPU memory inside task stderr will change.
That way we could establish what free memory needed on your host to start doing SETI MB on GPU.
Card with 256MB usually sufficient for CUDA MB.....
-
Switching of colors from 32-bit to 16-bit cure this problem. But with 16-bit one can forget about Aero effects so it seems the screensaver which can automatically switch 32-bit to 16-bit(and possiBly screen resolution) is needed.
Aero with all its effects can take quite a bit chunk of memory off of graphics cards, so all these 256MB cards might be bd`teb with aero turned off.
-
Switching of colors from 32-bit to 16-bit cure this problem. But with 16-bit one can forget about Aero effects so it seems the screensaver which can automatically switch 32-bit to 16-bit(and possibly screen resolution) is needed.
Aero with all its effects can take quite a bit chunk of memory off of graphics cards, so all these 256MB cards might be better with aero turned off.
I've figured as much, it would be aero, but i've run cuda with aero on before(with older drivers). I dont mind not being able to use the GPU on that computer. Just found it interesting that it would happen after a driver upgrade.
Oh well it was worth a try. Eventually I might get around to doing something about it. But I dont think I will remember to keep turning down the colours and loweing my resolution(1920x1200 lol) everytime I have to leave my computer for an extended period of time. I dont use the screensaver.
Well thanks for your time guys. This is really helpful. I'll keep these in mind if I get bored and want to mess around with something ;D .
-
Maybe new driver (or new CUDA lib - do you use CUDA 2.2 now ? ) need more GPU memory than older version...
Could you post memory values from stderr with and w/o Aero and with lower screen resolution?
(still interesting what mem amount is enough to successfully run CUDA MB now).
-
I've tried it with the new CUDA 2.2, tried with 2.1 and 2.0 and also stock and none of them ran. Also tried using different apps: random optimized ones and the stock one. But I have my computer without aero on and the resolution lowered. Just waiting to get some tasks. Then I'll let you know how they turn out.
Edit:
Running now. Heres the info you wanted.
setiathome_CUDA: CUDA Device 1 specified, checking...
Device 1: GeForce 8600M GT is okay
SETI@home using CUDA accelerated device GeForce 8600M GT
V10 modification by Raistmer
Priority of worker thread rised successfully
Priority of process adjusted successfully
Total GPU memory 268435456 free GPU memory 229027840
setiathome_enhanced 6.02 Visual Studio/Microsoft C++
Build features: Non-graphics VLAR autokill enabled FFTW x86
CPUID: Intel(R) Core(TM)2 Duo CPU T9300 @ 2.50GHz
Cache: L1=64K L2=6144K
-
And no CPU fall back? Fine, thanks
-
If this helps any, after I relogged into my computer it changed my resolution back to max. Then the current WU got a computation error. So I think the resolution is playing the biggest part in limiting.
This is from the stderr for the new one that started.
setiathome_CUDA: CUDA Device 1 specified, checking...
Device 1: GeForce 8600M GT is okay
SETI@home using CUDA accelerated device GeForce 8600M GT
V10 modification by Raistmer
Priority of worker thread rised successfully
Priority of process adjusted successfully
Total GPU memory 268435456 free GPU memory 223932416
setiathome_enhanced 6.02 Visual Studio/Microsoft C++
Build features: Non-graphics VLAR autokill enabled FFTW x86
CPUID: Intel(R) Core(TM)2 Duo CPU T9300 @ 2.50GHz
Cache: L1=64K L2=6144K
CPU features: FPU TSC PAE CMPXCHG8B APIC SYSENTER MTRR CMOV/CCMP MMX FXSAVE/FXRSTOR SSE SSE2 HT SSE3
libboinc: 6.4.5
Work Unit Info:
...............
WU true angle range is : 0.448219
Cuda error 'cudaMalloc((void**) &dev_t_funct_cache' in file 'd:/BTR/SETI6/SETI_MB_CUDA/client/cuda/cudaAcc_pulsefind.cu' in line 663 : out of memory.
setiathome_CUDA: CUDA runtime ERROR in device memory allocation (Step 3 of 3). Falling back to HOST CPU processing...
-
So "free GPU memory 229027840" worked, but "free GPU memory 223932416" didn't. The first is 218.42 MiB, the second 213.56 MiB. That may not be the only factor, but seems likely to be primary. Other data indicated the allocations take exactly 89 MiB, the critical point may be that 213.56 - 89 is less than 128 MiB, but 218 - 89 is more. Could there be a restriction that CUDA apps must leave 128 MiB free?
Joe
-
Seti does require more than 128MB. Its why my previous card(8400M GS) wouldnt work.
But might also need to leave some for the card to fuction properly. This is interesting. 128 is a lot though for a dedicated a card.