The blog from Valve seems to show that the GPU sometimes didn't do anything - which is what slowed down things.
That may not be the right interpretation of what we wrote. There were moments where either the CPU could be busy *or* the GPU could be busy, instead of having them concurrently active.
I'm hard pressed to think of situations where all parts of the system go idle (perhaps when waiting for files to load and the disk takes time to respond).
The issues touched on in the article have more to do with concurrency and keeping multiple workers (one GPU and two CPU threads) doing useful work consistently - i.e. 2 or 3 active instead of 1.