Spoke too soon, just got another crash, classic message

[drm:amdgpu_job_timedout [amdgpu]] *ERROR* ring sdma0 timeout, last signaled
seq=2423708, last emitted seq=2423710

On amdgpu.vm_update_mode=3 I got 3 days (about 12h per day) + 2/3h today
without crashing.