Clouds V3 Millions of Voxel with 2 suns

[KlausK]

Hi, after quite some cloud rendering tests popped up here lately I thought I`d try to find how many millions of voxel my system can handle.

Started with the Default of 10 Millions of Voxels (MoVs), went up to 100, then 500 MoVs and 1000 MoVs.

RAM usage maxed out at 99 percent with occasional paging to disk (which did not really affect rendering time because it was less than 10%)
after setting the value to 1800 Millions of Voxels.

I have no idea if that is a absolute number or if it depends on the scenery / clouds one sets up.
But I do know now that I will raise this setting in small steps from now on - simply doubling the number is not the way to go .
Had to wait quit some time before I could access TG again, hehe.

The file is attached for anyone interested.

CHeers, Klaus

[zaxxon]

Nice scene to experiment with. I've done a few similar tests with the voxel count with similar results. Personally, I've settled at somewhere between 150 to 500 million voxels depending upon the scene. While lower voxel counts essentially display the same imagery, to me the extra voxels add a crisper definition and allow some more detail in the darker areas. The results justify the extra render time to my eye, but some might disagree...  I'm running dual Zeon 5690's @ 3.47 ghz with 96 gigs of RAM and have 'pushed' the voxel count to 1,000 million (1000MoVs), but really couldn't see the additional improvement, and the RAM seemed adequate to the task. The V3 clouds with decent voxel counts is one of the main attractions to continue to use TG, simply the best clouds in the CG world in my opinion!

[RichTwo]

I wouldn't know what a voxel was or what it does if one bit me on my...  But I do like the underlighting for sure!

[pokoy]

If I remember correctly, the rule of thumb in my tests was 10 x MoV = 2 x Voxel density, which means that to achieve a really different/better result you need to increase MoV quite a lot. On really large cloud layers it's not really feasible due to how much RAM voxels take.

[WAS]

With dense clouds like this you're not really getting any benefit to more MoV. Take a look at the differences from your first example to your last. The  noticeable differences are in sliver lining and god-ray/glow shapes.

No offense, but it is a great example of how to optimize for the sake of RAM. xD

[KlausK]

@zaxxon: I agree. V3 clouds are really beautiful - I still have to learn how to "art-direct" them.
Since I am not doing this for a living render time is secondary - once the final render is set up it takes as long as it takes.
Leaves time for other hobbies

@RichTwo: Thanks! Me too.

@pokoy: I'll keep that rule of thumb in mind. True, RAM gets eaten pretty quickly by MoVs.

@WAS: No offense taken. As I wrote I wanted to find out how much Voxels I can squeeze into my RAM before it starts to page to hdd and make the scene unworkable.
On the other hand I think the three pictures can be used to demonstrate how much or how little influence the voxel count can have in this particular scene.
That's why the scene file is there as well to play with.

CHeers, Klaus

[WAS]

Indeed. Even the difference seems to be really at the pixel noise level,. Not really even greatly different overall shapes, they seem the same at second look. It's just the noise that's different, tad slightly larger silver lighting on the first image maybe.

One thing I've noticed, and just noted elsewhere, is v3 Voxels are tiny. They're very easy for windows to move from HDD to RAM so forth. I was rendering v3 cloudscapes on a 2gb netbook for a great while, and again on a Intel Compute Stick. The performance impact is mainly marginable, even well above 10% because what is actually because called back and forth is tiny really. Objects are where id get immediate crashes in just populating, because the objects can't be split up in RAM/HDD, they have to be moved around as a whole chunk. Right now on 8gb I get a long freeze if I were to punch in 500 MoV or so, but after it's done with it's freeze, everything operates as normal, Preview/RTP/Renders.

I'm assuming the same is said for polygonal caching as well. Matt's new robust sampler seems to greatly help with these scenarios as well and I see consistently faster renders with lower overall usage. Being back on 8gb of RAM again, it's actually a great help. Lol With objects I still run into trouble and currently limited to smaller sizes, as if they start paging, everything gets terribly slow due to wait time.