I've got Quad @ 3ghz and GTX 275 and I set my app_info to use 0.25 gpu so my gpu is working for all 4 cores. It is 50% load and 500mb used, so still room to spare. I just used the app_info posted above in plaintext and edited the gpu number. If everyone did this we would run out of workunits lol, but it would be more efficient. What will we do when we have caught up with the backlog I wonder? We are rapidly overtaking it; according to my math 109 days until very limited work.
I'm pretty sure you already are, its just you have a faster GPU, and maybe your CPU isn't quite as fast as mine, since mine is overclocked along That's a cuda project that is highly optimized.
I hope so, when I had six tasks running I noticed that instead of a typical completion time for a gpu unit of 1 1/4 hrs, I was getting 2-2.5 hrs for 3 wu's and estimated times of 4-6 for the remaining 3 wu's - they were at between 12-20% after 1 1/4 hrs, so I didn't think it was working OK..
Other people have reported times increasing, but not by taht much, Im suspecting that my cpu is a massive bottleneck and its not worth doing for me?
What cpu do you have and at what speed? Also, doing more than 4 at a time will massively slow it down since you only have 4 cores and they any extra have no cpu to run on and compete for time slots and bandwith.
Try with just 4 threads and see if it is more than a few percent slower.
the current Einstein CUDA app has such a low GPU utilization because the CPU is still doing a major amount of the work. Only one computationally expensive subroutine (FFT) is done on the GPU. So in order to feed a GPU you need an entire CPU core. Since you've got 4 cores (and they're not going to do any more work, even if you launch 12 CPU clients), you could feed 4 GPU tasks at a time. And you've got 4 GPUs, you on average you're running 1.3 GPU apps on your Fermis. That's why utilization is still low. You could take 2 of them out and see a reasonably higher value, though still far below 50% (going by the numbers others posted here).
Edit: and it doesn't help that an Athlon X4 810 is not exactly the fastest Einstein cruncher.
What you'll want to do is to fill the remaining time slices of your GPUs with other projects (hint: GPU-Grid can always use a couple more Fermis ;).
LHC and Rosetta are cpu-only, while nVidias can not use most of their power at Milkyway (using anything but ATIs there is a waste of electricity and computing resources). And SETI.. well, they probably like your Fermis, too :)
Yeah, and you can easily run 2, 3 or 4 tasks at once, 2 on a 460, 3 on a 470 and 4 on a 480, but even an 460 can run 4 MB tasks, time increases a little, depending on the A.R., though.
Haven't tried on at MilkyWay :).
An ATI 4850 can do a lot work there, ~40 -60K RAC, if one is very interrested in performance and who isn't, efficiency increases too,
Even now I run 2 ABP2 CUDA WU's, 2 on a 470 & 2 on a 480 .
GPU load hardly increases and doesn't get near 10% !
Compair a task done with a CPU and one done with a CPU & GPU, difference is very little maybe 20%, ofcoarse better then nothing.
It takes time and knowledge to write and compile a GPU app.
RE: I changed to 1 Are
)
Are you speaking of
Then there should be only four tasks running if you have four CPUs (using two GPUs in total).
Did you restart BOINC after the change?
Gruß,
Gundolf
Computer sind nicht alles im Leben. (Kleiner Scherz)
I've got Quad @ 3ghz and GTX
)
I've got Quad @ 3ghz and GTX 275 and I set my app_info to use 0.25 gpu so my gpu is working for all 4 cores. It is 50% load and 500mb used, so still room to spare. I just used the app_info posted above in plaintext and edited the gpu number. If everyone did this we would run out of workunits lol, but it would be more efficient. What will we do when we have caught up with the backlog I wonder? We are rapidly overtaking it; according to my math 109 days until very limited work.
Yes I was speaking of the
)
Yes I was speaking of the value
editing cuda devices
1 to 0.25 gives me 12 tasks running with gpu usage still at 6-9%
confused
There will be some serious timeslicing as I only have 4 x cpu!!!
What values do I need to change to get more than one unit at the same time on my gpu?
I'm pretty sure you already
)
I'm pretty sure you already are, its just you have a faster GPU, and maybe your CPU isn't quite as fast as mine, since mine is overclocked along That's a cuda project that is highly optimized.
I hope so, when I had six
)
I hope so, when I had six tasks running I noticed that instead of a typical completion time for a gpu unit of 1 1/4 hrs, I was getting 2-2.5 hrs for 3 wu's and estimated times of 4-6 for the remaining 3 wu's - they were at between 12-20% after 1 1/4 hrs, so I didn't think it was working OK..
Other people have reported times increasing, but not by taht much, Im suspecting that my cpu is a massive bottleneck and its not worth doing for me?
What cpu do you have and at
)
What cpu do you have and at what speed? Also, doing more than 4 at a time will massively slow it down since you only have 4 cores and they any extra have no cpu to run on and compete for time slots and bandwith.
Try with just 4 threads and see if it is more than a few percent slower.
MadMaC, the current
)
MadMaC,
the current Einstein CUDA app has such a low GPU utilization because the CPU is still doing a major amount of the work. Only one computationally expensive subroutine (FFT) is done on the GPU. So in order to feed a GPU you need an entire CPU core. Since you've got 4 cores (and they're not going to do any more work, even if you launch 12 CPU clients), you could feed 4 GPU tasks at a time. And you've got 4 GPUs, you on average you're running 1.3 GPU apps on your Fermis. That's why utilization is still low. You could take 2 of them out and see a reasonably higher value, though still far below 50% (going by the numbers others posted here).
Edit: and it doesn't help that an Athlon X4 810 is not exactly the fastest Einstein cruncher.
What you'll want to do is to fill the remaining time slices of your GPUs with other projects (hint: GPU-Grid can always use a couple more Fermis ;).
MrS
Scanning for our furry friends since Jan 2002
OK, fair enough, thanks for
)
OK, fair enough, thanks for the explanations and help guys
As for backup projects, I have seti, MW@home, lhc@home and rosetta! Im sure I can keep my rigs warm with that lot..
LHC and Rosetta are cpu-only,
)
LHC and Rosetta are cpu-only, while nVidias can not use most of their power at Milkyway (using anything but ATIs there is a waste of electricity and computing resources). And SETI.. well, they probably like your Fermis, too :)
MrS
Scanning for our furry friends since Jan 2002
Yeah, and you can easily run
)
Yeah, and you can easily run 2, 3 or 4 tasks at once, 2 on a 460, 3 on a 470 and 4 on a 480, but even an 460 can run 4 MB tasks, time increases a little, depending on the A.R., though.
Haven't tried on at MilkyWay :).
An ATI 4850 can do a lot work there, ~40 -60K RAC, if one is very interrested in performance and who isn't, efficiency increases too,
Even now I run 2 ABP2 CUDA WU's, 2 on a 470 & 2 on a 480 .
GPU load hardly increases and doesn't get near 10% !
Compair a task done with a CPU and one done with a CPU & GPU, difference is very little maybe 20%, ofcoarse better then nothing.
It takes time and knowledge to write and compile a GPU app.