BOINC ATI support thread... Unleashing the DRAGONs! Get your GPU's full worth here!

Ha, so evidently there is some issue between the way the current client, and the Ubuntu version of the openCL drivers, are identifying compatible openCL GPUs. This looks exactly like the issue that [email protected] has with openCL on AMD.

I don't have this issue on my systems, but they are all running Arch. These are the portions of the AMDGPU-Pro Ubuntu driver that I use:

• opencl-amdgpu-pro-icd_driverVersion_amd64.deb
• opencl-orca-amdgpu-pro-icd_driverVersion_amd64.deb
• libdrm-amdgpu-amdgpu1_2.4.95-driverVersion_amd64.deb

In addition to those packages from the Ubuntu driver the only other things needed is to get the openCL libraries added/linked in the correct locations, the standard libdrm from Mesa, and the Open source OpenCL ICD loader (ocl-icd). Here is a link to the package for opencl-amd from the AUR on Arch. https://www.archlinux.org/packages/e...86_64/ocl-icd/ You can see what it is bringing in from the Ubuntu driver if you click "View PKGBUILD" under "Package Actions".



I haven't tried it yet, but it replicating this should be fairly straight forward on Debian based distros without using all of the additional cruft from the AMDGPU-Pro stack.

Heh, ironically, I had already tried everything you said. I had manually linked all the libraries because the paths weren't correct (x86_64-linux-gnu instead of lib64, I think). Installing opencl-amd didn't work for me either, regardless which kernel I tried it on. The ICD loader wasn't necessary for my last try which got it working.

I'm not sure where else to post it but the RX 5700 XT properly undervolted with the GDDR6 overclocked is more energy efficient than the Radeon VII at hashing ETC. I'm wondering how to benchmark it in BOINC as I'm contributing to the [email protected] project in the Overclock.net team. I can see how much credit I'm receiving but I don't know what to compare it to.

Radeon VII is really a bigger pro card so I can see how the 5700s are more efficient.

My RX 580 can do 2x in 19-20min, 1080Ti can do a task in ~7:30, 1070Ti in ~10:50 and 1070 in 12:30.

Some 5700 comments here including some from @tictic;
https://einsteinathome.org/content/all-things-navi-10

Some people have posted Radeon VII numbers here:
https://einsteinathome.org/content/a...on-vii-vega-20

Do you allow stats export from [email protected] site? I don't see your name:
https://stats3.free-dc.org/stats.php...eah&team=10797

I didn't realize I had to change any options to make that visible.

[email protected] and some other projects follow GDPR regulations.

I get that but it wasn't very clear from an end user perspective that while I could see my own contributions, nobody else could. It appears that changing my settings has all my credit showing up TODAY when in fact it covers more than a week.

Yep, I saw your stats update today. For new users, everything will show up as it's like you're a new user today. There are still being that have been crunching for years that enable external stats export. I believe external sites wipe out the user tables so even if you crunched before the option started one would no longer display. Enabling the option would output all the credit as if done in one day.

All user stats can still be seen at [email protected] website. The option is for external stats export.

I'm still rocking my i7-6700K so I have it set to run 4 threads whenever CPU usage is below 45% total. That is pretty much all the time except for stalls here and there since I am not playing my more CPU intensive games at the moment. My RX 5700 XT A.E. is on a Wattman profile that runs it at ~100W total but still hashes at almost the full power of the card. The best power curve for this card is well below the factory specs so basically think of them all even reference as factory overclocked at this point. I'm hoping that the future RX 5800/5900 can do the same but at a less gimped core clock speed.

BTW it turns out the OpenCL drivers have a very small/silly bug that make Navi cards produce invalid results with this project. Until there's a fix I am leaving the GPU portion disabled.