If I understand correctly, you are able to run the v1.0 application with just a single task per gpu and it doesn't error out?
If so that is a new datapoint for troubleshooting the application on Ryzen systems and 8GB cards.
Correct. 1 Task per GPU with version 1.0 works fine so far across 4 Ubuntu 22.04 systems (3950x, 5900x, 2700x) with a mix of Pascal (1070ti) and Turing (1660ti, 2060, 2060 super) on driver 510.73.05.
With 2 Tasks per GPU on the 8GB cards (1070ti & 2060s) using version 1.0 the second task fails almost immediately.
Just for the fun of it and since I've a few GT710/720 cards lying around, I tested the new app (v0.95) and compare to the current production app (v1.28) on the same PC to see how fast or slow this 192-cuda core cards can do. Both draw about 9W of power. The v0.95 apps (11,248 sec, validated) is about 36% faster than v1.28 (17,542 sec. awaiting validation). The v1.28 task got validated and about 25% slower than comet-lake iGPU. Both cards have GK208B gpu processor and run with cuda 3.5 and opencl 3.0.
I'm a bit late to the party, but I just gave both 0.95 and 1.0 a try on my sole surviving NV GPU: a GTX 1080 Ti. Proprietary NV driver 510.73, OpenCL 3.0. Unfortunately, the tasks terminate with error after a few seconds and stderr lists signal 4 which I believe is 'invalid instruction'.
I believe I've set-up the anonymous platform correctly: EAH_SLEEP and *.alt files in the boinc-client directory, app_info.xml and binary (with execute permissions set) in the einstein.phys.uwm.edu projects directory.
Host is using an Intel Ivy Bridge-E CPU, which is supposed to support first-generation AVX - might an older CPU be an issue?
yes this is the issue. this was discovered more recently. Ivy Bridge CPU is a no-go.
needs to be something more modern, with AVX2 support.
but can you link to the host in question? In your hosts list, I only see one host with a Linux/Nvidia combination, but no errors reported on that host.
also, if you intend to run the AMD card as well in this host, you will need to add the appropriate AMD/ATI sections to the app_info.xml file, otherwise it will not run anything on the AMD card. running anonymous platform is all or nothing. you can't have some apps on AP and other apps standard.
Keith Myers wrote:If I
)
Correct. 1 Task per GPU with version 1.0 works fine so far across 4 Ubuntu 22.04 systems (3950x, 5900x, 2700x) with a mix of Pascal (1070ti) and Turing (1660ti, 2060, 2060 super) on driver 510.73.05.
With 2 Tasks per GPU on the 8GB cards (1070ti & 2060s) using version 1.0 the second task fails almost immediately.
This requires Linux. there
)
That really sucks there is no Windows version as not all of us run Linux due to situations beyond our control.
Of course, it would be nice if the E@H devs looked into this and just implemented it for everyone :)
This requires Linux. there
)
That really sucks there is no Windows version as not all of us run Linux due to situations beyond our control.
Of course, it would be nice if the E@H devs looked into this and just implemented it for everyone :)
You could try running a Linux
)
You could try running a Linux VM under windows. With GPU passed through to the VM. But I don’t know the specifics of doing that or if it’s possible.
the EAH devs are unlikely to do anything more now. There’s only about 2 months of work remaining.
_________________________________________________________________________
Just for the fun of it and
)
Just for the fun of it and since I've a few GT710/720 cards lying around, I tested the new app (v0.95) and compare to the current production app (v1.28) on the same PC to see how fast or slow this 192-cuda core cards can do. Both draw about 9W of power. The v0.95 apps (11,248 sec, validated) is about 36% faster than v1.28 (17,542 sec. awaiting validation). The v1.28 task got validated and about 25% slower than comet-lake iGPU. Both cards have GK208B gpu processor and run with cuda 3.5 and opencl 3.0.
GT720 1GB v1.28
GT710 2GB v0.95
PS: I should have run the v1.28 task on GT710 2GB card for better comparison but just been lazy. Ian informed me that I need 2GB to run v0.95.
11,248 seconds is actually
)
11,248 seconds is actually about 55% faster than 17,542 sec :)
to think of it another way, compare the work you could do per day for the two apps. there are 86400 sec in a day
86400 / 11248 = ~7.68 tasks/day
86400 / 17542 = ~4.925 tasks/day
7.68/4.925 = 1.559 times as productive, or 55.9% faster :)
still a great speed boost even for the slowest Kepler GPUs
it's too bad that the project only seems to have Intel GPU apps for Windows. I would like to try on my Xe GPU. but no app for Linux :(
_________________________________________________________________________
I'm a bit late to the party,
)
I'm a bit late to the party, but I just gave both 0.95 and 1.0 a try on my sole surviving NV GPU: a GTX 1080 Ti. Proprietary NV driver 510.73, OpenCL 3.0. Unfortunately, the tasks terminate with error after a few seconds and stderr lists signal 4 which I believe is 'invalid instruction'.
I believe I've set-up the anonymous platform correctly: EAH_SLEEP and *.alt files in the boinc-client directory, app_info.xml and binary (with execute permissions set) in the einstein.phys.uwm.edu projects directory.
I notice Keith is also running GTX 1080 Tis, also on Ubuntu 20.04, and seems to be running just fine.
https://einsteinathome.org/host/12600970
Host is using an Intel Ivy Bridge-E CPU, which is supposed to support first-generation AVX - might an older CPU be an issue?
I recognise the work for this application is running out, but there's still nearly a couple of months to go, according to the server status estimates.
Soli Deo Gloria
Wedge009 wrote: Host is
)
yes this is the issue. this was discovered more recently. Ivy Bridge CPU is a no-go.
needs to be something more modern, with AVX2 support.
but can you link to the host in question? In your hosts list, I only see one host with a Linux/Nvidia combination, but no errors reported on that host.
also, if you intend to run the AMD card as well in this host, you will need to add the appropriate AMD/ATI sections to the app_info.xml file, otherwise it will not run anything on the AMD card. running anonymous platform is all or nothing. you can't have some apps on AP and other apps standard.
_________________________________________________________________________
That's the one - no errors
)
That's the one - no errors because I reverted to standard application before retrying the tasks. It wasn't clear that AVX2 was a prerequisite.
And yes, of course, I know about anonymous platform from SETI@home days.
Soli Deo Gloria
If you can move the 1080Ti to
)
If you can move the 1080Ti to your threadripper system, it should run.
_________________________________________________________________________