Ram Usage

[mib2berlin]

During the "Triangle contest" tread testing i´ll try to get higher and higher with my polys.
The limit here was not my Vram, it was my Ram. During voxelizing the Ram usage goes up to 9 GB and crash octane or my system.
I have 4 GB Ram and 5 GB swap space on my system.
If i reduce the polys a little in my blender file, it worked and i saw that octane uses only 700 of my 895 MB Vram.
I want to buy a new 2GB gtx 460, do i need 20 GB Ram then?

I don´t think so, maybe this problem is only with linux?
In roubal´s latest thread he uses 1229.6 MB Vram with only 8 GB of Ram.

Hm, any sugestions, mib

[matej]

I want to buy a new 2GB gtx 460, do i need 20 GB Ram then?

Well, probably not, cause most of the vram will go for textures... and Octane will eventually support instances.
Do you really need all that polys for your work?

[mib2berlin]

Not often, the scene has 2.7 million faces (quads) and uses 470 MB in blender.
But this is not so much.
This is more a technical question why octane use so many memory during voxeliziation. (< I don´t know how to spell it)

Thank you, mib

[matej]

The guys also said that Octane as it is now it's not optimized. So we can probably expect better performance in this field. Just speculating...

[pixelrush]

It would be good to manage the ram use a little differently in the future to use less if thats possible.
Or maybe have an option to use a less demanding profile for those with less ram even if it takes somewhat longer to do?

Probably in usual use 6gb ram for 1gb vram is ok, 8 for 1.5, 12 for 2, so maybe 16 for 3, 24 for 4, 32 for 6? Not sure how it scales. I guess someone has to try it to find out.
Many people will get along fine with 6,8,12gb which is within reach and power users will need 16,24,32 - doesnt seem too unreasonable although it is a lot I agree.

I think some of the reason people will want a big vram is for a large image resolution.
For A0 prints and such that could take up 1gb space but it wouldnt be loaded so thats maybe 4gb ram saved.

If you can afford 6gb Tesla then you can afford 32gb ram I think if that whats required.
With 32gb I think you need a Xeon mobo though and it might be hard or impossible presently to find one with say 3x8x pci-e ?? Tyan, Supermicro, Asus, Intel seem to be the manufacturers but these are really for server so often have integrated graphics so that might be a gotcha.
I guess 24gb is the upper bound atm for Octane.
Given that, if the max memory use could be reduced by 25% it might be a good thing all round.

Edit: Even if you wanted to fill 6gb vram with just polys thats going to be something like 27m tris and its going to be a problem dealing with a model like that in Blender or Maya etc anyway.

[ROUBAL]

Hi, I noticed recently on one of my ocs files that loading it in Octane pre 2.3 uses a bit less VRAM
than loading it in 2.2 :Around 40MB less of a 740MB scene, it is not much, but could be a clue that memory management is on the way of improvement...

[pixelrush]

Just as a general note I think there is some possibility Cuda is reserving some space for itself that doesnt show in the stats - possibly 15% of vram - have to check that some more...if it is then really it ought to be shown in the Device Manager graph as that so people know they dont have that much space. Maybe like the attach

Edit: Done some more testing with a ~4.5m poly scene in 1gb vram, no textures etc, was able to get ~2000x1750px before it blacked out. For ~4.8m poly about a 700x450 image was the max I could get.
There is a small bug with the reporting or freeing of memory on res change I think that makes it hard to gauge where the limit is but it seems cuda needs to have about 11% free space, maybe 12.5% would be a good figure to use in the graph?? Perhaps if the allocation goes over the 87.5% it could show the overlap as red to warn you.

Edit2: I might also mention that when you get the vram really full it seems to styme zooming and panning the render image (ie with Ctrl) in the viewport. Perhaps it should be possible to specify the initial viewport zoom as 12.5,25,50,100% in the Preview Resolution pane so it can be set along with the res ahead of importing the mesh. This should centre the image in the viewport as well. For instance if you set a 8000x5000 image you would choose an initial zoom of 12.5% to be a handy size to monitor progress and you dont need to adjust anything once it was rendering.