Here's test 1:
At the top centre is the benchmark which has been rendered without RTA and with 24 atmo and 220 cloud samples which is equivalent to a detail setting of 0.76.
I then tested with 4 different atmo sample settings and 3 different cloud sample settings, all with RTA enabled.
The AA setting was 4 for all tests.
My conclusion in regard to atmosphere:
In this particular test adding more cloud samples did increase rendertime just slightly for all 3 cloud sample settings.
So, lowering atmosphere samples to <8 with RTA enabled does not change rendertimes dramatically.
This fits perfectly with my observation I told here before where my image with RTA @ AA8 and 8 atmo samples rendered 4 times faster when without RTA and 24 atmo samples.
My conclusion in regard to clouds:
In this particular test the noise levels are visible at the bottom left and right of the clouds.
When comparing the RTA OFF vs RTA ON conditions you'll notice that @ roughly 100 samples the result is similar, regardless of atmosphere samples.
However, rendertimes are ~2x longer.
Rendering with 40 atmosphere samples gives worse results and not much improvement in regard to rendertime.
Combined conclusion:
Rendering with RTA at similarly resulting noise levels is ~2x slower.
Rendering with RTA with more resulting noise is a little bit faster.
As discussed before, reducing atmosphere sample levels below 8 is not the cause for slow rendertimes with RTA.
So I will not advocate for RTA here anymore
and especially not when using AA>4, since that gives even bigger discrepancies as I explained before.
Due to the exponential sampling nature of AA this makes sense to me.
One can of course argue about the test setup, but so far these findings fit with what I've recently been experiencing. (after some discussions with Frank who mentioned it first to me that RTA isn't faster.)
Cheers,
Martin