That´s why I said it is not necessary with so many bounces, and it´s sort of Insane to do that 
unless having a double GPU installation with the latest best cards, and a lot of time and a decent payed project, or a renderfarm for animation.
the quality difference is negliable, just look at my samples, a slight difference…but not much, and as you can see from my previous posts with many different WDA results, they look good enough in my opinion, and they are very well lit and not silvery dark, and I´ve only used 40-50 volume bounces, on a machine that is an Asus Rog GTX 1080, bought in 2017 I think, so by todays standard the computer is outdated.
There are some specific light path outputs that needs to be plugged in the “right” volume inputs, that is part of the trick I use to get it this decent, without having to use thousands of bounces.
With cyclesX and rendering as I do, I get a decent feedback on the interactive renderer within a couple of seconds, and a decent refine that isn´t too noisy within some minutes, to clear up all noise you obviously have to set a lot of samples, the samples rendered in viewport arount 200 samples only, no noise reduction at all in post nor in viewports.
Keep in mind though, you said you tried with fractal noise on clouds with principle volume or volume scatter, that is a different thing VS VDB, the volume material on any mesh has No limit set by a voxel size, or the grid resolution…which means you can set whatever step rate you want…as low as 0,001 if you want.
But as I understand it…using a VDB file wich has a predefined voxel size resolution, it should not matter if you go lower that size in the step rate, it won´t make it look better, procedural noise density driven volume material is however Infinite in resolution.
So it´s a difference between a volumetric material on mesh, and VDB files, where VDB is limited and fractal volume density material isn´t
So Don´t try it with procedural clouds, use this asset from WDA instead.
Granted…the one with the backscatter has only the quarter resolution, while some others in here had the full resolution at over 2.73 gigabyt in filesize, don´t have the exact voxel resolution data in front of me now though.
I love the VDB rendering in blender, it´s decently fast, and give nice results, unlike Lightwave where it frankly is very bad in performance speed and lacks multiple volume bounces, so that is a big bummer.
On the other hand, for procedural texturing of clouds …I love lightwave, where blender is quite frankly, bad it what noise textures you can use, it baffles me that we can´t even access the Cloud fractal that is there for displacements, surfacing and volume displacement, but we can´t use it within the node material setup for volume density.
And the others are too limited, and it has to be setup by a huge mess of nodes which slows it down much more when it is evaluated.
With Lightwave procedurals, it can be done with just one or two to get nicer cloud fractals, but as mentioned…I can´t stay with lightwave much longer due to severe lacks in the rendering system, and lack of development.
I still have to evaluate the planetary cloud and atmosphere model in blender, which you may have something to say about
I just lack procedurals in blender.
Samples from Lightwave, real world scale diameter of the earth, volumetric clouds, and volumetric atmosphere, the atmosphere is foremost only tweaked by eyeballing scattering and absorbtion colors and assymetry (as it is called there and not anisotrophy( and density and shadow intensity on the volume item, the difference here between blender and lightwave here, is that you use a null and then apply volumetric handler on it, so it´s not a volumeric material on a mesh, you do not need a mesh for it to work, it has shadow density, which is how dense the shadow is when raycast on to ground, and you have shadow intensity which is the darkness control for the shadow within the volume material itself.
and it´s very easy to setup fractal textures with texture amplitude and density controls, however as mentioned before, it lacks true full multiple volume bounce scattering, you may trick your way around that a bit when usin fractals and curves fed in to the various scattering channels.
Feather type clouds…
Basic turbulence clouds…
Godrays are doable in there as well…quality is just not refined good enough, so abit noisy and I choosed to not use GI on the clouds for the test, due to long rendertimes with also a global scattering item for the godrays.
So I aim to try and find my way on to work with procedural fractals and volumetric materials for cloud layers as well in blender, but I find myself limited with the poor set of fractals that are there.
