The main problem I have with those systems is that if a bunch of WUs are downloaded with a short deadline I have see where the active WU will be suspended and the one just downloaded will be started. This can happen even if the one just downloaded has a deadline a few seconds earlier than the one just suspended. This is not unique to 6.6.20 as I've seen it occur with earlier BOINC releases as well.
This problem is fixed with BOINC v6.6.23 and later.
Most cannot run BOINC as a service and have it recognize CUDA, a limitation of the drivers in Vista (not sure if it will ever be addressed because it is a security thing).
6.6.20 has a couple severe bugs that seem to impact CUDA tasks more but can affect CPU tasks. If you want to run something past 6.5.0, then don't use 6.6.20, use 6.6.23...
6.6.24 and beyond can reject the second CUDA card/core for memory size problems. There is a flag in 6.6.25 that fixes that (cc_config)...
as to the unstable task scheduling, well, that is actually a LONG standing issue with BOINC that only because REALLY painful with CUDA and systems getting "wider". The more processors you have the less stable BOINC is/was.
The good news is that 6.6.28 seems to have fixed this (I first reported this in 2005, so I think it is about time)... the problem is that 6.6.28 has not been out that long so I cannot say it is better than 6.5.0 ...
Note, all 6.6.x versions from at least .15 through .28 all have a bug in the calculation of long term debt ... which may require periodic resetting of debt to keep your pending queue filled properly.
Hope these notes help ... for the moment, I still suggest 6.5.0 unless you are brave and foolhardy like me ...
did you install the latest NVIDIA drivers?
Michael
I reinstalled all the CUDA software and did a new installation but now with version 2.2. After a reboot the system has two CUDA devices found. I don’t know why the 2.0 drivers did not find my cards. Now it’s waiting for the E@H CUDA app.
did you install the latest NVIDIA drivers?
Michael
I reinstalled all the CUDA software and did a new installation but now with version 2.2. After a reboot the system has two CUDA devices found. I don’t know why the 2.0 drivers did not find my cards. Now it’s waiting for the E@H CUDA app.
Greetings,
TJ
While waiting you can try gpugrid@home project, I'm running it on my two compatible cards, till (if) einstein@home makes the units for it.
While waiting you can try gpugrid@home project, I'm running it on my two compatible cards, till (if) einstein@home makes the units for it.
Hi Ed,
I did a few WU's there indead during the last outage of E@H. However I noticed that the outflow air of the fan of the graphics cards became very hot, so the GPU should be even hotter. Installed software from nVidia to see its core temperature is not working on that machine, I have contacted them.
In the meantime I stopped with it, moreover it is curently very warm outside and even warmer in the room where the computers are.
Anyway, thanks for your tip.
While waiting you can try gpugrid@home project, I'm running it on my two compatible cards, till (if) einstein@home makes the units for it.
Hi Ed,
I did a few WU's there indead during the last outage of E@H. However I noticed that the outflow air of the fan of the graphics cards became very hot, so the GPU should be even hotter. Installed software from nVidia to see its core temperature is not working on that machine, I have contacted them.
In the meantime I stopped with it, moreover it is curently very warm outside and even warmer in the room where the computers are.
Anyway, thanks for your tip.
Yes thats a feature of the nvidia cards. Running hot that is :-)
The nvidia driver we seem to like over at GPUgrid is 182.50 at the moment. Their science app seems to have problems with later drivers. It also works with the Seti cuda multibeam app. I haven't tried other projects cuda apps.
I am currently running BOINC 6.6.36 on my farm. All my machines except one have cuda-capable cards installed. Apart from the debt issue that Paul mentioned it seems fairly reliable.
The exhaust from a GPU will be hotter than that of a CPU and the chip itself hotter as well. They're designed to do so because the lack of space makes cooling them harder. Normal full load operating temp for a GPU is ~80-100C vs 40-70C for a CPU.
RE: The main problem I have
)
This problem is fixed with BOINC v6.6.23 and later.
Most cannot run BOINC as a
)
Most cannot run BOINC as a service and have it recognize CUDA, a limitation of the drivers in Vista (not sure if it will ever be addressed because it is a security thing).
6.6.20 has a couple severe bugs that seem to impact CUDA tasks more but can affect CPU tasks. If you want to run something past 6.5.0, then don't use 6.6.20, use 6.6.23...
6.6.24 and beyond can reject the second CUDA card/core for memory size problems. There is a flag in 6.6.25 that fixes that (cc_config)...
as to the unstable task scheduling, well, that is actually a LONG standing issue with BOINC that only because REALLY painful with CUDA and systems getting "wider". The more processors you have the less stable BOINC is/was.
The good news is that 6.6.28 seems to have fixed this (I first reported this in 2005, so I think it is about time)... the problem is that 6.6.28 has not been out that long so I cannot say it is better than 6.5.0 ...
Note, all 6.6.x versions from at least .15 through .28 all have a bug in the calculation of long term debt ... which may require periodic resetting of debt to keep your pending queue filled properly.
Hope these notes help ... for the moment, I still suggest 6.5.0 unless you are brave and foolhardy like me ...
RE: did you install the
)
I reinstalled all the CUDA software and did a new installation but now with version 2.2. After a reboot the system has two CUDA devices found. I don’t know why the 2.0 drivers did not find my cards. Now it’s waiting for the E@H CUDA app.
Greetings,
TJ
Greetings from
TJ
RE: RE: did you install
)
While waiting you can try gpugrid@home project, I'm running it on my two compatible cards, till (if) einstein@home makes the units for it.
RE: While waiting you can
)
Hi Ed,
I did a few WU's there indead during the last outage of E@H. However I noticed that the outflow air of the fan of the graphics cards became very hot, so the GPU should be even hotter. Installed software from nVidia to see its core temperature is not working on that machine, I have contacted them.
In the meantime I stopped with it, moreover it is curently very warm outside and even warmer in the room where the computers are.
Anyway, thanks for your tip.
Greetings from
TJ
RE: RE: While waiting you
)
Yes thats a feature of the nvidia cards. Running hot that is :-)
The nvidia driver we seem to like over at GPUgrid is 182.50 at the moment. Their science app seems to have problems with later drivers. It also works with the Seti cuda multibeam app. I haven't tried other projects cuda apps.
I am currently running BOINC 6.6.36 on my farm. All my machines except one have cuda-capable cards installed. Apart from the debt issue that Paul mentioned it seems fairly reliable.
BOINC blog
The exhaust from a GPU will
)
The exhaust from a GPU will be hotter than that of a CPU and the chip itself hotter as well. They're designed to do so because the lack of space makes cooling them harder. Normal full load operating temp for a GPU is ~80-100C vs 40-70C for a CPU.