(In reply to comment #0)
> dpm was also manually enabled with 3.12 and before, so it's likely not
> related to that.

I think it's still worth testing if the problem occurs with DPM explicitly
disabled via radeon.dpm=0.

(In reply to comment #2)
> No, that would take forever and I have don't that much time ;)

Please reconsider. For every gaming session, you should be able to declare at
least one commit good or bad. It shouldn't take more than a dozen or so tests
to isolate the problem with git bisect.