Since we had a new binary for CUDA processing, I thought I'd run a few WUs to see if the situation had improved. Also running 3.05 WUs on the second cpu for the system.
Results:
Currently processing 1 CUDA WU and one cpu only WU. Est. time to completion for cpu WU is 7hrs, 1 min, 9 seconds (they're both almost complete). Est. time to completion for CUDA WU is 7hrs, 53 seconds.
Two previous CUDA WU were logged as taking 25,277.44/25,388.73 secs each.
Three 3.05 WUs that completed also took about the same time to process.
My opinion: the CUDA app (on an AMD X2 5200 system) is not viable presently because it takes both a cpu AND the gpu to process without any noticable improvement in throughput.
AND...if processing SETI on the same system, SETI will monopolize the gpu and not allow E@H to run.
If I understand correctly then the current run is a cuda assisted cpu run. The gpu is used for some calculations but not all. Right now your are mainly limited by the cpu. The gpu gives you just a boost of ca. 20-50%.
If your Q9300 is not overclocked then every 4GHz Core2 runs faster than your cpu with cuda help.
When there is a standalone app for the gpu then yours will fly ;)
The CPU is needed to feed the GPU from time to time and the GPU sends it back. So the CPU can be seen as a "handler" in the process. So for now there will be always a bit of help from the CPU (until they find other ways to handle).
I have crunching a CUDA WU with a GTX 285 in 5,5 hours. My wingmen has crunched the same WU with a CPU in 6 hours.
This is not a big difference .......
Hello,
I run the einstein cuda-app on a i7 with a GTX285.
It takes about 4.7-5 hours to complete, while I have seen wingman with faster CPU's (non-cuda) and they take 7-8.3 hours to complete. So I see a difference.
I have crunching a CUDA WU with a GTX 285 in 5,5 hours. My wingmen has crunched the same WU with a CPU in 6 hours.
This is not a big difference .......
Hello,
I run the einstein cuda-app on a i7 with a GTX285.
It takes about 4.7-5 hours to complete, while I have seen wingman with faster CPU's (non-cuda) and they take 7-8.3 hours to complete. So I see a difference.
I crunched a 3.10 using my CPU only, and found that compared to using the GPU it was about 80% as fast. In other words, the GPU crunches 20% faster than CPU by itself.
As others have commented, the GPU has a lot of potential I believe to become much faster while using less of the CPU. :)
Since we had a new binary for CUDA processing, I thought I'd run a few WUs to see if the situation had improved. Also running 3.05 WUs on the second cpu for the system.
Results:
Currently processing 1 CUDA WU and one cpu only WU. Est. time to completion for cpu WU is 7hrs, 1 min, 9 seconds (they're both almost complete). Est. time to completion for CUDA WU is 7hrs, 53 seconds.
Two previous CUDA WU were logged as taking 25,277.44/25,388.73 secs each.
Three 3.05 WUs that completed also took about the same time to process.
I've changed my mind on the following...comparing apples/oranges.
Quote:
My opinion: the CUDA app (on an AMD X2 5200 system) is not viable presently because it takes both a cpu AND the gpu to process without any noticable improvement in throughput.
The above is false because I'm trying to compare E@H WUs to APB1 WUs. I tried searching around for some non-CUDA APB1 WUs, but they've all rolled out of my stats. The rest of this post is valid though.
Quote:
AND...if processing SETI on the same system, SETI will monopolize the gpu and not allow E@H to run.
I crunched a 3.10 using my CPU only, and found that compared to using the GPU it was about 80% as fast. In other words, the GPU crunches 20% faster than CPU by itself.
As others have commented, the GPU has a lot of potential I believe to become much faster while using less of the CPU. :)
For a while there it seemed that the collective BOINC CUDA channeling was working. It seemed that the GPU was only using a small fraction of the CPU. But that seems to have dissipated for the time being with E@h. It seems we have a ways to go before we get to significant throughput increases of the magnitude we had hoped. Like others here though, I suspect it might take a while but we will eventually see some very significant improvements in throughput. It's slow going, but we'll get there. :-)
i downloaded all as posteted, installed it and started crunching.
Threre are two things i want to report:
fist, the usage of the gpu is low, it reaches only 62 deg (cent.), other applications (gpugrid, if it does not hang, or Seti) heat it to 74 deg.
Three times it caused a complete system hang just when finishing the wu at the beginning of the upload.
Vista64, E8400, 6GB, GTX260 and a lot of troubles.
The previous days the cuda-app. ran perfect and takes about 4 hours to complete on my i7 with a GTX285. The temperature of the GPU is very low and I like that. Good job.
Today BOINC has downloaded 6 cuda-wu’s, however one can be run at a time. So 7 of the CPU’s are doing now crunching for Einstein and that I do not like. It should be fine when they run 4 to 6 non cuda ABP and/or S5R5 WU’s. Can that be programmed?
This app has started fine on my 9800 GT. Thanks for developing an app compatible with 64-bit Vista/Win7! Other specs: Q9550, BOINC 6.6.36, Driver 190.38, v2.3 dll's.
Runtime projection is ~5 hours compared to 5h 45min with the 3.09 CPU app. So it's in effect slower since the GPU resource is used. First task 137298125.
Following a work request for both CPU and CUDA work I got both but also the 4-hour deferral. The scheduler log is not logical:
It sent S5R5 work despite not finding an app for it? Or maybe it wanted an S5R5 CUDA entry. My app_info includes 3.05 for S5R5, 3.09 for CPU ABP1 and 3.10 for CUDA ABP1.
I've found how to run 4 S5R5 tasks while working on ABP1 on CUDA device.
It's simple. Just freeze all the ABP tasks, then restart BOINC. And after that unfreeze one of ABP tasks and have fun ;)
It seems to run smoothly (but before that I killed some WU's with old driver). Will keep eye on it and report after a while.
Since we had a new binary for
)
Since we had a new binary for CUDA processing, I thought I'd run a few WUs to see if the situation had improved. Also running 3.05 WUs on the second cpu for the system.
Results:
Currently processing 1 CUDA WU and one cpu only WU. Est. time to completion for cpu WU is 7hrs, 1 min, 9 seconds (they're both almost complete). Est. time to completion for CUDA WU is 7hrs, 53 seconds.
Two previous CUDA WU were logged as taking 25,277.44/25,388.73 secs each.
Three 3.05 WUs that completed also took about the same time to process.
My opinion: the CUDA app (on an AMD X2 5200 system) is not viable presently because it takes both a cpu AND the gpu to process without any noticable improvement in throughput.
AND...if processing SETI on the same system, SETI will monopolize the gpu and not allow E@H to run.
Seti Classic Final Total: 11446 WU.
RE: If I understand
)
The CPU is needed to feed the GPU from time to time and the GPU sends it back. So the CPU can be seen as a "handler" in the process. So for now there will be always a bit of help from the CPU (until they find other ways to handle).
Greetings from
TJ
RE: I have crunching a CUDA
)
Hello,
I run the einstein cuda-app on a i7 with a GTX285.
It takes about 4.7-5 hours to complete, while I have seen wingman with faster CPU's (non-cuda) and they take 7-8.3 hours to complete. So I see a difference.
Greetings from
TJ
RE: RE: I have crunching
)
I crunched a 3.10 using my CPU only, and found that compared to using the GPU it was about 80% as fast. In other words, the GPU crunches 20% faster than CPU by itself.
As others have commented, the GPU has a lot of potential I believe to become much faster while using less of the CPU. :)
RE: Since we had a new
)
I've changed my mind on the following...comparing apples/oranges.
The above is false because I'm trying to compare E@H WUs to APB1 WUs. I tried searching around for some non-CUDA APB1 WUs, but they've all rolled out of my stats. The rest of this post is valid though.
Seti Classic Final Total: 11446 WU.
RE: I crunched a 3.10 using
)
For a while there it seemed that the collective BOINC CUDA channeling was working. It seemed that the GPU was only using a small fraction of the CPU. But that seems to have dissipated for the time being with E@h. It seems we have a ways to go before we get to significant throughput increases of the magnitude we had hoped. Like others here though, I suspect it might take a while but we will eventually see some very significant improvements in throughput. It's slow going, but we'll get there. :-)
(Click for detailed stats)
Hi, i downloaded all as
)
Hi,
i downloaded all as posteted, installed it and started crunching.
Threre are two things i want to report:
fist, the usage of the gpu is low, it reaches only 62 deg (cent.), other applications (gpugrid, if it does not hang, or Seti) heat it to 74 deg.
Three times it caused a complete system hang just when finishing the wu at the beginning of the upload.
Vista64, E8400, 6GB, GTX260 and a lot of troubles.
Regards,
Alexander
To the developers, The
)
To the developers,
The previous days the cuda-app. ran perfect and takes about 4 hours to complete on my i7 with a GTX285. The temperature of the GPU is very low and I like that. Good job.
Today BOINC has downloaded 6 cuda-wu’s, however one can be run at a time. So 7 of the CPU’s are doing now crunching for Einstein and that I do not like. It should be fine when they run 4 to 6 non cuda ABP and/or S5R5 WU’s. Can that be programmed?
Thanks and success.
Greetings from
TJ
This app has started fine on
)
This app has started fine on my 9800 GT. Thanks for developing an app compatible with 64-bit Vista/Win7! Other specs: Q9550, BOINC 6.6.36, Driver 190.38, v2.3 dll's.
Runtime projection is ~5 hours compared to 5h 45min with the 3.09 CPU app. So it's in effect slower since the GPU resource is used. First task 137298125.
Following a work request for both CPU and CUDA work I got both but also the 4-hour deferral. The scheduler log is not logical:
It sent S5R5 work despite not finding an app for it? Or maybe it wanted an S5R5 CUDA entry. My app_info includes 3.05 for S5R5, 3.09 for CPU ABP1 and 3.10 for CUDA ABP1.
Edit - link to scheduler log, spelling
I've found how to run 4 S5R5
)
I've found how to run 4 S5R5 tasks while working on ABP1 on CUDA device.
It's simple. Just freeze all the ABP tasks, then restart BOINC. And after that unfreeze one of ABP tasks and have fun ;)
It seems to run smoothly (but before that I killed some WU's with old driver). Will keep eye on it and report after a while.