Out-of-Core Memory Paging

[Chris]

Is Out-of-Core Memory Paging being considered to avoid GPU RAM restriction?

Cheers,

Chris

[pixelrush]

For accessing more textures you mean? or more work space in general?
My uneducated thought is that Octane requires all it needs for rendering to be resident in vram.
The number of textures is limited by hardware addressing I think.
I know that as cuda/hardware progresses more instructions are available including ram <-pcie-> vram and vram <- pcie -> vram. Gpu computing is still opening up.
Probably Refractive will make good use of relevant things as they come along but most likely it wont be for some time as they are concentrating on other things like the list of features promised in v1.
After that people will be clamoring for still more shaders and UI improvements and.... The list of things to do goes on more hours than they have available.
Bear in mind each cuda update and now the plugins are extra work as well.
If core code needed to be rewritten to take advantage of something there would need to be a really good reason to do that immediately.
My guess is that unbiased on gpu has a long and exciting future ahead of it.
Keep watching this space and upgrading in good faith

[t_3]

even if paging would be possible (don't know), it'll degrade performance a lot, as the maximum bandwidth over pcie is a fraction of the vram memory bandwidth (~192gb/s for a gtx 580 vs. 8gb/s for pcie2.0 x16).

just an example: lowering the memory speed of a gtx 580 from 2000 to 1500mhz let ms/sec drop from 3.96 to 0.68(!) in the benchmark scene. now if you had not only a cut by 25% but one of 95%, it'll be just unusable. saying that, i won't make a bet that we will see this anytime soon...

[pixelrush]

Yeah true. I see the pcie 4.0 spec is set to arrive about 2015 or so and that's probably 32gb/s. Somewhat less in reality I suppose. I'm not sure how graphics cards and cuda will be at the time thats available in a mobo. 3-4x faster perhaps...there could well be creative ways to use memory if the transfer is fairly small and infrequent. More interactivity or experimentation with the scene for example.
If you think of using your mouse to pause your render and restart it immediately that's probably ~0.3 second, not a major deal for the user, yet quite a few MB could have moved in/out in that time.
Guess we'll find out how these capabilities can best be exploited in due course...

[Chris]

It would be for geometry i guess. You can find some info here:

http://raytracey.blogspot.com/2012/01/o ... eased.html It's about Nvidia's Optix which have the capability, released a couple of days back.

http://www.centileo.com/news.html Link to CentiLeo which also have this capability, watch the YouTube video

I'm not sure how it works, but it looks promising. It would be cool to get a word from the devs on this one

Cheers,

Chris

[pixelrush]

Interesting. Yeah where are those developers these days? Gone very quiet. Must be hard at work....
The plane model only seemed to have very simple materials though and it looked more like direct lighting. Didn't notice any textures other than that used for the background or really any shadows either.
I notice too although it said 3-10 frames/sec it actually seemed to take between 1-3 seconds to update the camera. Not so interactive as we are used to. I presume it is like backface culling to the camera view and the BVH is done on the fly.
Still at 360m poly that is enormous. Most PC would struggle to deal with a 12m poly model. Its hard to imagine a need for quite that size unless you are a plane maker with a major database and even then you would wonder if it is necessary to render that amount of detail at once.
Perhaps the concept would be handy for Octane archviz walk throughs but tuned to more like 48m polys + textures?
You would have to have some sort of way of calling and positioning a collection of objects from a library like using Blender empties as proxies or surely you would never handle setting up the scene...
Isn't Cycles based on Optix? Maybe Cycles get this in the future...
I guess Refractive will end up having to cover this development somehow or at least in a way that's useful to the typical Octane user.
Perhaps rather than a 'static mode' fed by cpu as now a 'progressive mode' could be set where a gpu can be dedicated to generating the BVH on the fly while the others render and they are updated every few seconds or so by pcie - like viewing a room at a time? Dunno
How many polys is enough? Is a walk through of a museum complete with a crowd necessary?
Be interesting to hear what other users might want from such a set up and how they might use it.