There are a couple things to look at here.
Most importantly your tests would be more conclusive if you had changed *fewer* parameters for each comparison. Changing *both* atmosphere and cloud sample levels doesn't give you an accurate comparison, especially for render time. I would either leave the atmosphere samples at 16 to 32 for *both* image tests for consistency, or use 128 for *both*, thus elminating it as a source of potential noise in both tests. That way the atmosphere sampling contributes equally to render time in both tests and you can really see the impact of just the increased cloud samples and the relative quality increases.
It's important to understand that increases in atmosphere sampling are *not* necessary to correspond with increased cloud sampling and, except in heavily shadowed scenes, complex lighting, or in the presence of atmospheric rays I never recommend more than 64 atmosphere samples. Sometimes it's necessary, but you should remember that these settings are separate for a reason (otherwise we'd have an overall "atmosphere samples" slider), and they should be used with discretion.
One of the reasons for the need of such high cloud samples in your particular examples is the combination of high density and a noisy input function. Obviously if that seems to be the only way to create the desired effect then yes you need to use high samples to get good quality, but to my eyes that level of noise/roughness is almost always unrealistic, and high density is not necessarily the only way to achieve a better look of "solidity". If you reduced the roughness of the input function and the density a bit you would need fewer samples for equivalent quality, and you'd probably increase overall visual realism as well.
- Oshyan