Since Einstein task list doesn't allow filtering by app...
Yes, it does. Use the drop-down list to the top-right of the main task list - you can select any app that you currently have tasks for in the list. You can also sort the list by sent time, to bring the current (v0.05) tests to the top.
I've completed my eyeball of the short BRP4 tasks on the UHD 620. At the time I scraped the results, the score was
All (844)
In progress (0)
Pending (85)
Validation inconclusive (89)
Valid (558)
Invalid (112)
Error (0)
I only looked at the 112 invalids. I didn't see a single Windows partner in the whole set - if I had any, they must be in one of the other buckets.
Of the invalids, 108 were validated by two ARM machines, and 4 by two aarch64 machines. 46 also had an invalid ARM machine in the mix. There were 4 invalid aarch64 tasks, and 4 validate errors.
I think Bernd is right: the validation problem is purely in the variability of the ARM results.
(copy of the full spreadsheet available on request)
even though the 390 drivers support this GPU for use as a display, it seems they don't support CUDA on this GPU anymore. a little digging showed that CC 2.1 (Fermi generation) lost CUDA support in CUDA SDK 9.0 and after, meaning an older version of the driver (and kernel for compatibility) will be necessary for any CUDA processing with these old GPUs. I will try Ubuntu 18.04 and the latest 340.108 drivers.
Ouch. Nvidia never were good with software. When I download software for an old AMD, I'm given the latest version that supported that card, not a newer version which half supports it.
AMD has far more software/driver issues. this is really me going against the grain to make such an old card work on an relatively modern OS. this era of drivers was REALLY picky about which linux kernel and compiler you had installed. and around this time, nvidia on linux was a headache to get working since there wasn't as much push to have the OS devs working on compatibility. this is all a thing of the past and modern drivers and modern linux kernels all work much more smoothly these days.
BUT, I was a bit incorrect about the nature of Fermi support anyway. While support was pulled from the CUDA toolkit, it *should* remain in the drivers
Fermi Architecture Support. Fermi architecture support is being deprecated in the CUDA 8.0 Toolkit, which will be the last toolkit release to support it. Future versions of the CUDA Toolkit will not support the architecture and are not guaranteed to work on that platform. Note that support for Fermi is being deprecated in the CUDA Toolkit but not in the driver. Applications compiled with CUDA 8.0 or older will continue to work on Fermi with newer NVIDIA drivers.
so one of two things are possible. these old linux drivers are just bunk (the only ones I've been able to successfully install are the ones packaged and distributed by the Ubuntu devs), or there really is something wrong with the hardware. this card will process FGRPB1G though (opencl app). but I don't have any other cuda apps to test on it.
I guess I'll just wait for the cuda55 Windows app to be released and give it another go there. Windows Nvidia drivers worked better in this era, so maybe the windows app will give different results.
but as of now, it's looking not very viable for old GPUs to run this app anyway, and if not, there's no reason to stick with such an old cuda version. that's really what I'm trying to test.
it appears this app requires at least compute capability 3.0, so Fermi and older cards are out anyway.
from my log (this line didn't show previously with the v0.03 app):
Quote:
2022-08-23 13:12:07.1267 [PID=17846] [version] Checking plan class 'BRP7-cuda55'
2022-08-23 13:12:07.1295 [PID=17846] [version] reading plan classes from file '/BOINC/projects/EinsteinAtHome/plan_class_spec.xml'
2022-08-23 13:12:07.1296 [PID=17846] [version] parsed project prefs setting 'gpu_util_brp': 1.000000
2022-08-23 13:12:07.1296 [PID=17846] [version] CUDA compute capability required min: 300, supplied: 201
this further proves my point that using such an old cuda version doesn't make sense, if it's only for the sake of compatibility. Compute capability 3.0 is supported all the way to CUDA 10. and it looks like drivers supporting CC 3.0 cards go all the way to CUDA 11.4
Bernd, if you updated this app to cuda100, you would keep the same compatibility and get a little speed boost for the more modern GPUs. but even going to the very latest CUDA 11.7 still supports CC 3.5.
Is that a technical limit, or an Einstein server deployment limit? I can't find my Fermi, but I've got a GTX 670 'Kepler' with CC 3.0
given that I was unable to process any tasks (instant errors) with my CC 2.1 Fermi card, I'd say the app is unable to process on anything less than 3.0. and they set the server limit accordingly to prevent sending to older GPUs.
but since CC 3.0 is still supported by much more modern drivers and cuda versions, there's no reason to stick with a cuda 5.5 just for the sake of compatibility if the app doesnt even support these old GPUs anyway.
even the v0.05 nvidia OpenCL app they just uploaded wont work on it. instant error. but this might be a problem with the application, need to test on something more modern. unless they are using some function that's not supported by this hardware. maybe they need to gatekeep anything less than opencl 1.2?
Peter Hucker of the Scottish
)
Yes, it does. Use the drop-down list to the top-right of the main task list - you can select any app that you currently have tasks for in the list. You can also sort the list by sent time, to bring the current (v0.05) tests to the top.
A shame I didn't close the
)
A shame I didn't close the wager ( mine is computer # 12941363 ) :
... so we await the tiebreaker !?
Cheers, Mike
( edit ) Given that they don't agree sufficiently well, it will be interesting to deduce why they differ ..... if we can that is.
( edit ) Oooops, wager outcome not satisfied - nothing yet sent out.
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
I've completed my eyeball of
)
I've completed my eyeball of the short BRP4 tasks on the UHD 620. At the time I scraped the results, the score was
All (844)
In progress (0)
Pending (85)
Validation inconclusive (89)
Valid (558)
Invalid (112)
Error (0)
I only looked at the 112 invalids. I didn't see a single Windows partner in the whole set - if I had any, they must be in one of the other buckets.
Of the invalids, 108 were validated by two ARM machines, and 4 by two aarch64 machines. 46 also had an invalid ARM machine in the mix. There were 4 invalid aarch64 tasks, and 4 validate errors.
I think Bernd is right: the validation problem is purely in the variability of the ARM results.
(copy of the full spreadsheet available on request)
Peter Hucker of the Scottish
)
AMD has far more software/driver issues. this is really me going against the grain to make such an old card work on an relatively modern OS. this era of drivers was REALLY picky about which linux kernel and compiler you had installed. and around this time, nvidia on linux was a headache to get working since there wasn't as much push to have the OS devs working on compatibility. this is all a thing of the past and modern drivers and modern linux kernels all work much more smoothly these days.
BUT, I was a bit incorrect about the nature of Fermi support anyway. While support was pulled from the CUDA toolkit, it *should* remain in the drivers
NVIDIA CUDA TOOLKIT 8.0.61 Release Notes, p.11
so one of two things are possible. these old linux drivers are just bunk (the only ones I've been able to successfully install are the ones packaged and distributed by the Ubuntu devs), or there really is something wrong with the hardware. this card will process FGRPB1G though (opencl app). but I don't have any other cuda apps to test on it.
I guess I'll just wait for the cuda55 Windows app to be released and give it another go there. Windows Nvidia drivers worked better in this era, so maybe the windows app will give different results.
but as of now, it's looking not very viable for old GPUs to run this app anyway, and if not, there's no reason to stick with such an old cuda version. that's really what I'm trying to test.
anyone else get a Fermi card working?
_________________________________________________________________________
v0.05 cuda55 app is about
)
v0.05 cuda55 app is about 20-25% slower than the v0.03 app.
v0.05, 3080Ti, 2x tasks, 811 seconds
v0.03 (via AP), 3080Ti, 2x tasks, 666 seconds
optimizations removed to improve validation?
there are also some occurrences of what appears to be "stuck" tasks. that are running but not progressing at the rate they should.
_________________________________________________________________________
Ian&Steve C. wrote: anyone
)
I think I've got an old 470 in working order - it's in the parts bin now, but I could mount it somewhere, or lend it to someone.
It's the one I bought for SETI Beta message 39386 (HTTP only, no certificates) - 15 May 2010.
AH, I think I've solved
)
AH, I think I've solved it.
it appears this app requires at least compute capability 3.0, so Fermi and older cards are out anyway.
from my log (this line didn't show previously with the v0.03 app):
this further proves my point that using such an old cuda version doesn't make sense, if it's only for the sake of compatibility. Compute capability 3.0 is supported all the way to CUDA 10. and it looks like drivers supporting CC 3.0 cards go all the way to CUDA 11.4
Bernd, if you updated this app to cuda100, you would keep the same compatibility and get a little speed boost for the more modern GPUs. but even going to the very latest CUDA 11.7 still supports CC 3.5.
_________________________________________________________________________
Is that a technical limit, or
)
Is that a technical limit, or an Einstein server deployment limit? I can't find my Fermi, but I've got a GTX 670 'Kepler' with CC 3.0
Richard Haselgrove wrote:Is
)
given that I was unable to process any tasks (instant errors) with my CC 2.1 Fermi card, I'd say the app is unable to process on anything less than 3.0. and they set the server limit accordingly to prevent sending to older GPUs.
but since CC 3.0 is still supported by much more modern drivers and cuda versions, there's no reason to stick with a cuda 5.5 just for the sake of compatibility if the app doesnt even support these old GPUs anyway.
your Kepler card will probably work .
_________________________________________________________________________
even the v0.05 nvidia OpenCL
)
even the v0.05 nvidia OpenCL app they just uploaded wont work on it. instant error. but this might be a problem with the application, need to test on something more modern. unless they are using some function that's not supported by this hardware. maybe they need to gatekeep anything less than opencl 1.2?
https://einsteinathome.org/task/1341629858
_________________________________________________________________________