(In reply to Chris Wilson from comment #11)
> An idle gpu is always reported as running at the min(idle) freq. You have to
> be very careful in the workload you construct that it does indeed keep the
> gpu busy if you want to use a sampling method.
> 

Apparently the test case uses queries, not the sampled method of metrics
measurement.

> No sign of a bug yet...

No disagreement there, however any further suggestion as to what to look at to
figure out what is going wrong is welcome.