but this might be a problem with the application, need to test on something more modern.
well there doesnt seem to be anything inherently wrong with the v0.05 nvidia-opencl application. running just fine on a GTX 1060 6GB.
it seems Fermi is still too old, even for the opencl app. this GPU only supports OpenCL 1.1
Bernd, you might need to put a opencl check in the scheduler to not send to GPUs with less than OpenCL 1.2. or recompile the app (or another version of the app) that works with older GPUs with opencl 1.1 if you want to support these old cards.
I've run about 40 of the current v0.05 MeerKAT tasks on three hosts, all running Windows on three different AMD Navi-class GPU cards.
I'm running about 10% computation errors. Mine generally occur after what appears to be full normal run times. This is hugely worse than the same machines have seen running 1.28 Gamma-ray pulsar binary search.
I've only got six validations. Every one is with a quorum partner who is also running Windows/AMD.
Two of those validations came on Work Units which had already piled up more than ten errors while computing on other people's machines. Unlike mine, those compute errors generally had come quickly (8 to 63 elapsed seconds). They were generally on Windows machines running NVidia cards (so maybe that is why that application has vanished from the application list for the moment).
On the machine that has run the most tasks, I have clear indications that 2X beats 1X and that 3X beats 2X both in GPU utilization and in productivity indicated by task completion times. This result is likely card specific and perhaps task specific, as I think the GPU memory usage is pretty substantial.
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.
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.
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.
[[snip]]
anyone else get a Fermi card working?
I've got 12 Tahiti chips working on Windows 11 (10 year old AMDs). I've never had a software problem in the 27 years I've used AMD. No problems with Radio apart from when they physically explode. That reminds me I've got a capacitor to replace.
If this page takes an hour to load, reduce posts per page to 20 in your settings, then the tinpot 486 Einstein uses can handle it.
For the latest app version (0.05) we did change the precision of some variables from float to double (to increase validity and improve validation). For CUDA this required us to change the kernel calling API. This works well on Linux, but apparently doesn't on Windows. The OpenCL app obviously doesn't work either on NVidia, but thanks to the recent testing we now at least have a problem that we can reproduce deterministically on our own machines, and thus fix more easily. We're working on it. For now I disabled all app versions again.
Ian&Steve C. wrote: but this
)
well there doesnt seem to be anything inherently wrong with the v0.05 nvidia-opencl application. running just fine on a GTX 1060 6GB.
it seems Fermi is still too old, even for the opencl app. this GPU only supports OpenCL 1.1
Bernd, you might need to put a opencl check in the scheduler to not send to GPUs with less than OpenCL 1.2. or recompile the app (or another version of the app) that works with older GPUs with opencl 1.1 if you want to support these old cards.
_________________________________________________________________________
I have errors on all my 27
)
I have errors on all my 27 WUs, every wingman also have errors, except one which ran for almost 6000 secs. Seems something isn't going well.
The Nvidia 0.05 opencl win
)
The Nvidia 0.05 opencl win application fails both on GTX-970 and GTX-1070. Runtime is only 15-20 seconds.
I've run about 40 of the
)
I've run about 40 of the current v0.05 MeerKAT tasks on three hosts, all running Windows on three different AMD Navi-class GPU cards.
I'm running about 10% computation errors. Mine generally occur after what appears to be full normal run times. This is hugely worse than the same machines have seen running 1.28 Gamma-ray pulsar binary search.
I've only got six validations. Every one is with a quorum partner who is also running Windows/AMD.
Two of those validations came on Work Units which had already piled up more than ten errors while computing on other people's machines. Unlike mine, those compute errors generally had come quickly (8 to 63 elapsed seconds). They were generally on Windows machines running NVidia cards (so maybe that is why that application has vanished from the application list for the moment).
On the machine that has run the most tasks, I have clear indications that 2X beats 1X and that 3X beats 2X both in GPU utilization and in productivity indicated by task completion times. This result is likely card specific and perhaps task specific, as I think the GPU memory usage is pretty substantial.
There is a new Linux
)
There is a new Linux version (at least the time is updated).
0.05 (BRP7-opencl-nvidia) 23 Aug 2022, 13:46:01 UTC
It is running OK for me on a GTX 1060 after 20 minutes and 50% complete.
But it is set to use only 0.5C + 1NV, so I set it to use a full core.
Yeah the v0.05 Nvidia-OpenCL
)
Yeah the v0.05 Nvidia-OpenCL version seems to work fine (when the task doesn’t get stuck). But it’s about 20-25% slower still than the CUDA app.
_________________________________________________________________________
Richard Haselgrove
)
Call me stupid if you wish, but this is the main task list? https://einsteinathome.org/account/tasks
There's no drop down box displaying for me. Applications is just black text, not selectable in any way.
If this page takes an hour to load, reduce posts per page to 20 in your settings, then the tinpot 486 Einstein uses can handle it.
Peter Hucker of the Scottish
)
try this one, which is the same as the "Tasks" link at the bottom of your main Account Dashboard.
https://einsteinathome.org/account/tasks/0/0 (you're missing the /0/0)
_________________________________________________________________________
Ian&Steve C. wrote:AMD has
)
I've got 12 Tahiti chips working on Windows 11 (10 year old AMDs). I've never had a software problem in the 27 years I've used AMD. No problems with Radio apart from when they physically explode. That reminds me I've got a capacitor to replace.
If this page takes an hour to load, reduce posts per page to 20 in your settings, then the tinpot 486 Einstein uses can handle it.
Short report:We did
)
Short report:
For the latest app version (0.05) we did change the precision of some variables from float to double (to increase validity and improve validation). For CUDA this required us to change the kernel calling API. This works well on Linux, but apparently doesn't on Windows. The OpenCL app obviously doesn't work either on NVidia, but thanks to the recent testing we now at least have a problem that we can reproduce deterministically on our own machines, and thus fix more easily. We're working on it. For now I disabled all app versions again.
BM