Here's a tip that may help if, like me, you're running other projects as well.
I currently run 4 cuda tasks on each of my gtx 580s. I've found the time to complete 4 can jump from about 90 minutes to over 120 minutes if there's too many cores being used for other crunching. However I'd like these cores to be used when einstein is not running cuda tasks.
The solution? Increasing from .2 to .5 for the brp cuda tasks. This way when they run the core count used for cpu tasks drops by 2 for each gpu used. When they're finished and non-einstein gpu tasks are running then those cores are available again for normal crunching.
Well having just bought and installed a new NVidia based card - not OC'ed but I could have - on this machine last week, I am thus doing CUDA when I wuzn't before. I can see the reason for the excitement here : ~ 1800 seconds per WU for 500 credits. I stuffed around a bit getting the right driver ( 8.17.12.6666 ) - oddly 'Windows Update' didn't pick up any need for that - and it just purrs along like a happy cat now with no 'Error while computing'. Now I reckon I'll just continue to leave things as per spec - whatever BOINC/E@H throw at it, no config file fiddling etc - to remain a 'fiducial example' if you like. I'll run the CUDA WU's in 'Use GPU always' local preference, and see if anything prangs, bogs down or overheats - watch a movie or two in Itunes and play a spot of Supreme Commander Forged Alliance perhaps.
Being a wide-blade/high-volume fanboy and copper-heatsink-by-the-kilogram type of guy, my case and usage keep the CPU temp to about 30 C. The MSI Afterburner utility reports around 332 MB in use, GPU temp of 43 C with (i) a BRP3cuda32 ( 0.2 CPUs + 1 NVidia GPUs ) running and (ii) one monitor only as a 1440 by 900 desktop.
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
to better utilise the cards peformance, ( as I do on a pair of GTX 460 OC 45 and 51 degrees as I type ! ) you need ro run 2 WU's as the time to complete each is about the same, but you'll use > 70% of the GPU's capacity.
to better utilise the cards peformance, ( as I do on a pair of GTX 460 OC 45 and 51 degrees as I type ! ) you need ro run 2 WU's as the time to complete each is about the same, but you'll use > 70% of the GPU's capacity.
IMHO.
dunx
Hi Dunx. Thanks, I was thinking of moving to that later, maybe when I work this stuff out better. I think 3 simultaneous WU's is a bit ambitious, and still retain system usability for my other purposes. I saw some cards in-store that had over 1G RAM but the price jump was ..... breathtaking.
Cheers, Mike.
( edit ) Flash of inspiration followed by vaulting ambition overleaping : I dropped the old card back in as well. But I had to rearrange who was in what PCI slot ( thus triggering Windows to reinvent the device listing ) for reasons of cooling and physical hindrance - both are quite long cards, and don't you hate how the mobo SATA ports are placed sometimes? So I've tripped up more than a couple ( err, about 3 dozen actually ... sorry dear wing-ladies&gentlemen ) of CUDA WU's because of that. Silly me. So I'll have to wait a day or so until BOINC trusts me again. Now that's a 1G card that isn't needed for display purposes ..... so he's the guy I'll experiment with multi-WU's on first up.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
Two Wu's should be fine I barely see 700 Mb used leaving a massive amount for even a 1080P display...
And it may be a lot easier to set-up 2 W.U.'s on both cards as I have... : - )
Due to "work in progress" I have a fourth PC in bits, otherwise I'd have my 8800 GTS 320Mb running in my i5 system ( display uses in-built graphics ! )
Does it ever end, I've seen a new GTX 480 for less than the 460 GTX OC now...
DOH !
dunx
P.S. All I have is cc_config plus an App_Info from elsewhere on this site...
Well I've dropped in the old old graphics card with 512MB also. After sliding in the cc_config.xml - thanks for that timely hint dunx - they're all chugging along :
notably with 8 HT cores still doing 8 non-CUDA tasks as well! So now I have to cook up an app_info ??? Probably can get 2 + 1 + 2 = 5 simultaneous CUDA's at least ..... ;-)
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
Well I've dropped in the old old graphics card with 512MB also. After sliding in the cc_config.xml - thanks for that timely hint dunx - they're all chugging along :
notably with 8 HT cores still doing 8 non-CUDA tasks as well! So now I have to cook up an app_info ??? Probably can get 2 + 1 + 2 = 5 simultaneous CUDA's at least ..... ;-)
Cheers, Mike.
Sorry, Mike, I'm afraid you're aiming a bit too high there. BOINC doesn't yet have the granularity to allow that sort of fine control, even with an app_info file. And they've had enough of a struggle even to get it to this state - I don't think the will is there to do it all over again, just at the moment. We're just in the middle of giving them a hard time over OpenCL...
Any figure you slide in to enable multiple tasks per GPU will apply to all GPUs (well, all NVidia GPUs) active at the time. You might do best by disabling that 9500GT (or transferring it to another host - you'll never run more than one task in 488 MB) and running two each on the other cards.
then why can't i run 3 WU in
)
then why can't i run 3 WU in 1024 mb?
At a guess, your OS needs
)
At a guess, your OS needs >300MB in order to manage your desktop.
Well ! I had some bother
)
Well !
I had some bother trying to run two GTX 460's in one PC... the WU times were very poor ( SLI ? ). So I built a 2nd i7 870 PC to split them up...
Tonight I decided to experiment with 3 x WU's per GPU.
1 WU = 38 minutes
3 WU = 105 minutes or 35 per WU ! i.e. FASTER !
2nd PC is taking 96 minutes as it's an OC version 5% faster.
Or 32 minutes per WU !
Chuffed !
dunx
P.S. Big thank you to Claggy for his valuable input !
Here's a tip that may help
)
Here's a tip that may help if, like me, you're running other projects as well.
I currently run 4 cuda tasks on each of my gtx 580s. I've found the time to complete 4 can jump from about 90 minutes to over 120 minutes if there's too many cores being used for other crunching. However I'd like these cores to be used when einstein is not running cuda tasks.
The solution? Increasing from .2 to .5 for the brp cuda tasks. This way when they run the core count used for cpu tasks drops by 2 for each gpu used. When they're finished and non-einstein gpu tasks are running then those cores are available again for normal crunching.
Well having just bought and
)
Well having just bought and installed a new NVidia based card - not OC'ed but I could have - on this machine last week, I am thus doing CUDA when I wuzn't before. I can see the reason for the excitement here : ~ 1800 seconds per WU for 500 credits. I stuffed around a bit getting the right driver ( 8.17.12.6666 ) - oddly 'Windows Update' didn't pick up any need for that - and it just purrs along like a happy cat now with no 'Error while computing'. Now I reckon I'll just continue to leave things as per spec - whatever BOINC/E@H throw at it, no config file fiddling etc - to remain a 'fiducial example' if you like. I'll run the CUDA WU's in 'Use GPU always' local preference, and see if anything prangs, bogs down or overheats - watch a movie or two in Itunes and play a spot of Supreme Commander Forged Alliance perhaps.
Being a wide-blade/high-volume fanboy and copper-heatsink-by-the-kilogram type of guy, my case and usage keep the CPU temp to about 30 C. The MSI Afterburner utility reports around 332 MB in use, GPU temp of 43 C with (i) a BRP3cuda32 ( 0.2 CPUs + 1 NVidia GPUs ) running and (ii) one monitor only as a 1440 by 900 desktop.
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
Hi Mike, to better utilise
)
Hi Mike,
to better utilise the cards peformance, ( as I do on a pair of GTX 460 OC 45 and 51 degrees as I type ! ) you need ro run 2 WU's as the time to complete each is about the same, but you'll use > 70% of the GPU's capacity.
IMHO.
dunx
RE: Hi Mike, to better
)
Hi Dunx. Thanks, I was thinking of moving to that later, maybe when I work this stuff out better. I think 3 simultaneous WU's is a bit ambitious, and still retain system usability for my other purposes. I saw some cards in-store that had over 1G RAM but the price jump was ..... breathtaking.
Cheers, Mike.
( edit ) Flash of inspiration followed by vaulting ambition overleaping : I dropped the old card back in as well. But I had to rearrange who was in what PCI slot ( thus triggering Windows to reinvent the device listing ) for reasons of cooling and physical hindrance - both are quite long cards, and don't you hate how the mobo SATA ports are placed sometimes? So I've tripped up more than a couple ( err, about 3 dozen actually ... sorry dear wing-ladies&gentlemen ) of CUDA WU's because of that. Silly me. So I'll have to wait a day or so until BOINC trusts me again. Now that's a 1G card that isn't needed for display purposes ..... so he's the guy I'll experiment with multi-WU's on first up.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
Two Wu's should be fine I
)
Two Wu's should be fine I barely see 700 Mb used leaving a massive amount for even a 1080P display...
And it may be a lot easier to set-up 2 W.U.'s on both cards as I have... : - )
Due to "work in progress" I have a fourth PC in bits, otherwise I'd have my 8800 GTS 320Mb running in my i5 system ( display uses in-built graphics ! )
Does it ever end, I've seen a new GTX 480 for less than the 460 GTX OC now...
DOH !
dunx
P.S. All I have is cc_config plus an App_Info from elsewhere on this site...
1
Well I've dropped in the old
)
Well I've dropped in the old old graphics card with 512MB also. After sliding in the cc_config.xml - thanks for that timely hint dunx - they're all chugging along :
notably with 8 HT cores still doing 8 non-CUDA tasks as well! So now I have to cook up an app_info ??? Probably can get 2 + 1 + 2 = 5 simultaneous CUDA's at least ..... ;-)
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
RE: Well I've dropped in
)
Sorry, Mike, I'm afraid you're aiming a bit too high there. BOINC doesn't yet have the granularity to allow that sort of fine control, even with an app_info file. And they've had enough of a struggle even to get it to this state - I don't think the will is there to do it all over again, just at the moment. We're just in the middle of giving them a hard time over OpenCL...
Any figure you slide in to enable multiple tasks per GPU will apply to all GPUs (well, all NVidia GPUs) active at the time. You might do best by disabling that 9500GT (or transferring it to another host - you'll never run more than one task in 488 MB) and running two each on the other cards.