(In reply to comment #24)
> Created attachment 101420 [details]
> dmesg with working bapm patch
> 
> I can confirm that the latter patch (attachment 101191 [details] [review] [review])
> solves the issue for me, I haven't tried the earlier patch because it didn't
> work for Collin. I can re-test that if that's still useful.
> 

Probably not necessary.

> Attaching dmesg; I compared the probed power states and they appear similar.
> What's bapm?

It allows the CPU and GPU to share TDP headroom.  E.g., if the CPU isn't busy,
the the GPU can user higher performance states longer and vice versa.