Memory management

[yamanash]

really doubt that. octane is for example about 5 times faster than luxrender on an i7 quad (roughly/depending on/etc.).
now make it 10 times slower because of uma...

Take a look at what these guys are working on: http://www.centileo.com/news.html Even barely using any vram they still can get a 400million poly scene with gi going at around 3 fps, id say that is significantly faster than an i7. Don't get me wrong I think you guys have a top rate product here and I will likely be purchasing it soon but for this type of technology to become a feasible solution for the wide range of people who use it, I think that the memory barrier needs to be addressed. Just my 2 cents lol. You guys are doing a great job so far though, keep it up!

[t_3]

Take a look at what these guys are working on: http://www.centileo.com/news.html Even barely using any vram they still can get a 400million poly scene with gi going at around 3 fps, id say that is significantly faster than an i7. Don't get me wrong I think you guys have a top rate product here and I will likely be purchasing it soon but for this type of technology to become a feasible solution for the wide range of people who use it, I think that the memory barrier needs to be addressed. Just my 2 cents lol. You guys are doing a great job so far though, keep it up!

be assured, i fight with memory and texture limits every other day. by the way i'm not associated with octane/refractive/otoy, only forum mod.

now this video is already 1yr old, well known here but you need to judge what you see... if you strip all advanced mat properties and use dl with simple gi in octane, you can have in fact another factor 10 speedup (even more). if you then use uma thus make it ten times slower again, it'll still be blazingly fast. but what's the point? no medium or transmission attributes, no advanced glossy or specular mats, don't ever think about complex mix materials, just no nothing. for large scale technical or scientific visualization this could be the way to go (and one of the still amzing facts is, that the guy managed to handle 470mio polys with only 16gb system ram), but octane is about photorealistic unbiased rendering, using complex material attributes. you just need to connect a transmission node to a diffuse mat, and you will see ms/sec dropping for each and every new transmission node you add to a scene.

i would love to see out of the core rendering with octane's quality but i doubt it would be possible to manage a competitive speed. with the last arion version for example they say "no more limits" but in fact they move rendering to the cpu if a scene doesn't fit into gpu vram. i don't think that random control would have chosen this approach, if it would have been possible to keep a good rendering speed while using uma.

still there might be some ways to use uma while keeping low render times, maybe using this technique in combination: http://www.youtube.com/watch?v=ZQ6TFZE5QEU (also already quoted here). at the moment you need only a tenth of the samples to achieve the same quality, a ten times slower rendering process might make sense. on the other hand cpu based pathtracers could also use it and in the ende they would be faster again

[yamanash]

What you say is true, but obviously using only vram to render on the gpu right now is not a feasible solution for people looking to do anything further than rendering a simple scene. I just think that adding the feature would be good just as a fall back because as we all know the amount of vram on a consumer level or even professional level gpus these days just doesn't make it feasible to do complex scenes for things like movie production and so on. I know nvidia and weta digital developed a gpu ray tracer that can page to system ram, I don't think they would have gone to all that trouble if it want worth implementing right? It too could be falling back to the cpu, I don't know. Do you know if that is something Octane might add in the future?

[t_3]

Do you know if that is something Octane might add in the future?

no, i don't ... but i'd like to know too

[a.boeglin]

I'm still insisting to keep the main goal here. Using system ram or other gpu's vram would benefit only to render larger scenes. We keep a full speed solution for small scenes, this point is primordial. But as said earlier, I think people don't really care having a final render slower if while setting things up they go 10 times faster than with other tools. And another point is that u need to optimise exchanges between vram and system ram. If you need to access system ram at each ray you cast its not going to work for sure and may be 10 times slower as you say, but if you do it every 1000 rays, or if you are able to cast rays while transfering datas you can get a heavy optimisation. You need to find a pattern that would work for this case. I don't say it is an easy task but there are solutions for sure.

[glimpse]

actually that's pretty interesting topic You started there =) let me get in too =)

yeah, everything what You say it's maybe logical, maybe it would be nice, but..i don't want that - you know why? =)

the reason is quite simple - there are tools for specific purposes =) if the tool is for everything it has to trade of something to be well rounded - do I want that? NO, 'cos i'd rather have good render engine, as fast as possible and do scenes that I could optimise for it - that's it.

If the scene becomes to large, You can always switch to Vray (or maxwell =) and if You dont like to get too deep into parameters Vray could be setuped for unbiath mode - basicly bruteforcing everything - with server board You can get 100Gb+ of ram if You want, into SRX from Evga You can stuff nearly 100 and with two high end OC'ed procesors with eight Cores in each You get enough power to render pretty fast to, but at what cost?

alternativelly, if You want to do the work with GPU - choose the GPU accelerated renderers like INDIGO =) Yeah, you get out of limitation but for that You'll sacrifise the speed. if I'm right they run on regular ram and CPU with GPU boost.

it's only few options to mentions. The scene is open & pretty sure soon we'll see biased renders on the scene optimised of at least accelerated by GPU's =)


Now Refractive team has chosen it's own path - to develop everything and stuff information on vRam - there's an reason for that. If You take HYBRID aproach there will be system bottlenecs - If You need those..as mentioned there are other products.

Personally I don't want any trade offs, so i'd better see FULL GPU engine - That's what I like about OCTANE. It's insanely fast! if You need a lot of ram - save Your money, look for some projects and invest a bit into Quadro/Teslas - with kepler architecture we should see upgraded models =) i wouldn't be surprised if 8-9gb models would appear in the end of this year.


Bottom line. It's not the the engine, nor the hardware is going to make the work for You. When You look around, whos on the scene =) guys with a lot of limitations shows things that are way better that from all those with hyper PC's and renderfarms in their disposal. to reinforce the idea with facts - Bertrand Benoit and his octane corners made with 1gb card (265gtx if I'm right =). Isn't it detail ENOUGH? dout the fact =) so what could be done with 6gb quadro vRAM? what we could do this engine when instancing will be rolled out..I could not image to be honest.. =)

there's a reason for everything. =) Don't waste Your time dreaming and even asking things that doubtly would ever be implimented. Weta digital has milions of dolars in their hands to develop tools, and these tools are used by people that do a lot themself - they don't just push the buttons, majority of them do code at least a bit.

Now What You want to get here is F1 speeds with the drive comfort of Bently.. =)

Make Your choise..Majority of Octane users take Octane for it's speed & ease of use, not for versality. It's very young render engine compared to it's competitors and it has bright future, but for now it is what it is.. =)

We have to appreciate these facts, be aware of limitations and simply move on =)



Seen really nice works from You in the forum recently =)

Push Yourself a bit further and let's wait for instancing! maybe thinking about a card or two with high amount of vRam is an option too =) but yeah, these are costly - i know that.


keep renderin'


Glimps

[bloomendale]

Speaking about archviz i don't get why you need to render all scene at once. All these trees can be (should be) added in post (so many great archviz companies do). There are plenty of real trees on the internet (royalty free, free...). Even if you need your 3d tree just render it alone with some matte-like mat to receive shadows and comp it in.
I think that 99% "don't fit in memory" problems can be solved this way.

[Mr_Nitro]

Hello,
I am recently trying to use Octane for some industrial renderings. The models are very heavy.. about 3-4 gb of data... as a potential customer I would say that I wouldn't mind waiting for the program to manage an automatic pc to gpu ram transfer times , during a render, ( I think like chunks of size equal to the gpu memory available).
As per the raw scene interaction, I would be ok to disable the interactive rendering part , use a conventional shaded triangles view , and translate the scene (in chunks , in necessary) at render time.
As a rendering platform I think the last limitation I would want is about size of scenes, doesn't matter if there
are slow downs, but I would want the render to finish no matter what.

my2cnts


mrn