The Scottish Boinc Team

Seems like a useful sort of a chap to know. It has done wonders for your WCG returns too

As to your question on AMD. Has any one looked at https://rocm.github.io/ROCmInstall.html ??

Seems like this is being used on the 'official' Docker BOINC images for AMD GPU crunching.

I am going to take a look at the whole BOINC in Docker setup, https://github.com/BOINC/boinc-client-docker, when I get home although I think all my nVidia & AMD cards are simply too old to test anything GPU wise.

I think my mate was looking at that but didn't get anywhere with it. Dirk might be in the know.

I've just put a pair of RX580 into a machine and I see one of them is showing 81C - is that too hot (not rising above that). The other one is 20C cooler but I suspect the hot one is the one without much gap to feed the fans. Case is open of course

When do they start missing cycles to prevent overheating?

My AMD usually run around 78-80C with 38% fan. I use a custom fan profile on Windows so that if they go up to around 82C the fan goes up to 50% and at 85C goes to 100%. They were bought in 2010 and ran flat out for 4 years at that and are running again now. So I would say from experience that is a normal temperature for them running Collatz.

Off the top of my head there are certain things that cause Milkyway failures.
Firstly the GPU must be Double Precision, some older GPU's and mobile GPU's are not.
OpenCL must work.
A lot of older chipped cards can not run the nbody tasks. I remember having to run all my 69xx series cards under old drivers or they don't work. Seperation tasks should work though.
Is collatz using openCl or Cuda ?