file size and gpu memory

[siliconbauhaus]

Having been burned with maxwell and never really been happy with fry, I'm looking over things here and I'm quite impressed with what I've seen so far. It would also be nice to be able to get away from max.

Right now I don't have a nvidia card so I can't try out the demo, but before I go out and lay down some dosh on a gtx I wanted to ask about the amount of ram on a card.

I currently render arch vis with mental ray and many scenes tend to use around 91 % of my 4gigs of ram. I know I could do with more ram ( I just need more work in to fund it ) but I'm unclear as to how Octane uses the GPU ram and what limitations it may cause? If I was to buy a card with say 1gb of ram on it, how complicated a model can it handle?

[[gk]]

The model can be pritty darn complicated yet swift in octane.
You would run into issues regarding textures and shaders with multi level high res bitmaps pritty fast I recon. You say it yourself, your scene sometime fills 91% of your 4gb.

Octane is isolated to the amount of memory located on the cards, so whatever amount you got, thats your cap.

2 cards cant share. so you wont gain anything on the amount stacking cards, only on render speed.

[siliconbauhaus]

I guess its hard to pin down the amount of memory needed. I can work in a file and it sits there around the 60% mark and only on a render does it climb to almost max.

I just tried a scene I was doing in mental ray and exported it out to fry. When I opened it with fry I was at 50% ram and it went to 55% when rendering ( I just used a grey diffuse on everything so no bitmaps) so perhaps ram isn't as big of an issue as I was thinking.

I really need a new vid card anyway, so I when I get one I'll give it a try.

[radiance]

I think if you get a GTX260 with 1.8GB Ram,
or wait for the GTX300 series you will be able to render some pretty complex scenes.

you can get away with about 100MB video ram per million polies.
say 8 million polys = 800MB
add to that 100MB for a large framebuffer (eg render resolution)
that will leave you 900 MB for textures.

if you load specular and bump maps as greyscale (in octane it's a 'float texture' instead of an 'image texture'),
and you're sparing with your texture resolutions,
you should be able to render some pretty complex scenes already.

sparing = don't load 4096x496 stock texture images for small objects or objects at large distances,
eg try to keep your texture's to the minimum resolution to make then render ok...

Radiance

[pixelrush]

Hard to get a word in these days.

There are 1792mb vram cards out there which will accomodate larger scenes that you should buy in preference to a 896 mb one in your case I think, but its likely Octane may not allow you to work with the same scenes you have now maxing out 4gb -or potentially even more.
Have a look in the gallery at some of the images done with <1gb to give you an idea.
Currently, if you can afford it, a Quadro 5800 (based on GTX285) has the most vram at 4gb. I cant say if this will give you an equivalent work space...
EDIT: btw MSRP for the FX5800 is US$3500 vs $300 for a GTX260/1792

Soon a new Tesla card based on Fermi will be out and this has the option of 3 or 6gb vram I believe, a maxed Quadro probably the same capacity, and I think we will see Geforce versions with 1.5 and later 3gb.
It says in Nvidia product blurb this generation of card will be 3-4 faster than GTX2xx for Octane type use...we'll see..
Tesla cards will be very many $$$$ though, quadro versions $$$, and Geforce versions also $$. I would think these latter ~$500

HTH

[siliconbauhaus]

Thanks for the info / advice.

I just checked the one scene I did and it's 2,540,447 worth of polys but it would probably be more if it included a couple of tree proxies in there.

It does look doable though.....just need to go get a card I guess.

[pixelrush]

If you think of your present situation as 4gb-1gb for O/S and applications = 3gb then a new fermi Geforce with 3gb for say ~$600 may yield much the same rendering work space...and *possibly* run say 36x faster than a quad cpu solution.

I guess for the same money you could get 2x GTX260/1792mb and run smaller scenes or less textures at say 24x speed..

btw nice renders on your site

[siliconbauhaus]

Well I'm looking into the details of what I'd have to put together to run this.

I don't think I can run two gpu's in my current hp box so I may have to build one ( I have a old boxx r2 looking for a new lease of life)

Thanks for the comment on the renders.

[pluMmet]

An nVidia x95 card has 2 GPUs on it. but does it share the RAM? ie. nVidia 295 with 1792 MB GDDR3. Is that 896 MB per GPU?

[radiance]

An nVidia x95 card has 2 GPUs on it. but does it share the RAM? ie. nVidia 295 with 1792 MB GDDR3. Is that 896 MB per GPU?

no, it's 896MB per gpu / scene maximum, it does'nt share it.

Radiance