moskitos anyone?

[fellazb]

Hi there,

A few days ago this engine caught my eye and I'm seriously wondering what the bottleneck is (except for having a massive workstation and running on a GPU environment I pressume).

http://www.cebas.com/?pid=productinfo&prd_id=175

The one thing that I find the most amazing feature is that it handles all sort of (Autodesk) shaders without converting them and is very smoothly integrated within the 3DS MAX environment. Has  anybody used this software already and are there some things that Corona could benefit from?

[crazyman]

meeeh.... no OpenCL only CUDA, and max 2 GPU's per machine. On the top of that, you MUST have CUDA device, no CPU render is supported. Also some features you expect them to be present, are not natively supported such as Fur and Hair. If I had to choose, I rather go for VrayRT.

[Ludvik Koutny]

I was actually quite interested in it...  but it failed even at rendering teapot on a plane... (i may have done something wrong, and probably did, but i did not want to spend much time figuring out how to do something as easy) so i gave up quickly :)

[fellazb]

So it's just another "too good to be true" thing then? Thanks for sharing your thoughts on this so far. Ah well, I wasn't planning on leaving Corona that soon :)

[Ludvik Koutny]

It looks really good on the paper. Marketing departments are always doing a great job of making something look great on website presentation or paper. I was actually very excited about it...  when i read they support everything from MentalRay (which i know) and that they have overcome most of the limitations. And that their interactive rendering is really seamless. But once i tried it, actual truth was very different.

 Lots of things are still not supported, or working in a weird manner, and the main limitation - GPU memory- which Redshift has found some way around, is present in moskito...  so unless you have at least Titan Z with 12GB RAM, there's not much of a chance to render some more complex scenes.

Usually, you see some well built web presentation and hype video, and then once you try yourself, you are disappointed. I like Corona because Ondra does the exact opposite. There's not much of an aggressive marketing trying to oversell the product, and once you give it a try, you are pleasantly surprised, because it is above your expectations, instead of below them.

[fellazb]

@Rawalanche

That's a very true statement you've put down there. This is why Corona for me is so addictive, because I didn't need to spend hours and hours to figure things out, but could quickly jump on board and devolop from there. (And a big hurray for the mat convertor)

Funny you mentioned that you know MR, of course you do :) Your tutorial on the Bekerman site is a classic for a lot of MR users. But it's even funnier when I discoverded it was actually you who put up an epic forum topic on the 3DS MAX beta where you stated the lack of MR development and it was so true. I enjoyed every piece of it (expect for you I guess) and I'm happy to see you back on a much more respectful forum.

Cheers

[Juraj]

so unless you have at least Titan Z with 12GB RAM, there's not much of a chance to render some more complex scenes.

TitanZ has 6GB, it's dual-chip unit. And that makes this renderer just as obsolete to me as every other one, until shared memory becomes available and proves to be efficient (which might not happen).

I wasn't that impressed here from start...seems odd way to choose. Maybe they're targeting very broad general market with those "all autodesk shaders" (because I don't see how that would really interest pro users).

[Ondra]

TitanZ has 6GB, it's dual-chip unit. And that makes this renderer just as obsolete to me as every other one, until shared memory becomes available and proves to be efficient (which might not happen).

Just a side-note here: modern CPUs are memory-bound, not FPU-bound. If somebody invents a way to get that much data over all the buses, then... I can make Corona MUCH faster :D

[Juraj]

I remember Octane dev saying the same few years ago. Might be reason why it's much faster than the mixed CPU-GPU solutions, which only show negligible speed improvement by adding GPU to mix.

So than we're stuck with 4000 euros costing Teslak40/Quadro6000, only cards with 12GB of vram. Really amazing...
Are most GPU-renderer customers even aware of this fact ? Or they just see "100x faster!" and insta-buy.

[Captain Obvious]

I don't know if you guys were aware, but Redshift happily renders "unlimited" texture data and hundreds of millions of triangles with limited amounts of VRAM. Its memory cycling seems pretty efficient. Frankly, it deals with complex scenes on a memory-limited GPU than Corona does on with oodles of system memory. The idea that GPU renderers have to be limited to just the size of the video memory isn't really the case any more, as Redshift demonstrates. It also uses irradiance caching, photon mapping and other biased methods. So GPU renderers don't need to be stuck in slow-as-molasses naive path tracing mode, either.

[lacilaci]

I remember Octane dev saying the same few years ago. Might be reason why it's much faster than the mixed CPU-GPU solutions, which only show negligible speed improvement by adding GPU to mix.

So than we're stuck with 4000 euros costing Teslak40/Quadro6000, only cards with 12GB of vram. Really amazing...
Are most GPU-renderer customers even aware of this fact ? Or they just see "100x faster!" and insta-buy.

When it comes to gpu rendering I believe that fans of "quick, but expensive" gpu rendering solution are mostly consisted of people that focus more on product visualization (from simple stuff to cars and other vehicles possibly) and smaller amount of interior viz.. So, it makes somewhat sense to go for "100x faster" gpus when you care only about the product model with an hdri which might be sometimes pretty far from reaching a 6GB VRAM limit... But anyways, so far I think that for "simple stuff" an expensive gpu is overkill and for complex scenes it is unusable if you don't want to spend ages on optimizing the scene to fit that VRAM...

[Juraj]

I don't know if you guys were aware, but Redshift happily renders "unlimited" texture data and hundreds of millions of triangles with limited amounts of VRAM. Its memory cycling seems pretty efficient. Frankly, it deals with complex scenes on a memory-limited GPU than Corona does on with oodles of system memory. The idea that GPU renderers have to be limited to just the size of the video memory isn't really the case any more, as Redshift demonstrates. It also uses irradiance caching, photon mapping and other biased methods. So GPU renderers don't need to be stuck in slow-as-molasses naive path tracing mode, either.

Can that be because of Embree in Corona ? Everytime I switch to Embree in Vray, I get almost zero speedup in complex scenes (:-D I dunno) but my memory doubles instantly.
I've heard even Octane can be pretty efficient in what it allocates during rendering compared to CPU renderers but still. Regarding the 6GB, I've heard from one car-specific studio that the more complex car models only start at 12GB alone, so it would have to be
really simple product visualization.

But I'll have a look at this Redshift, it got me quite interested when it was posted here as comparison. Esp. the IC.

[lacilaci]

I don't know if you guys were aware, but Redshift happily renders "unlimited" texture data and hundreds of millions of triangles with limited amounts of VRAM. Its memory cycling seems pretty efficient. Frankly, it deals with complex scenes on a memory-limited GPU than Corona does on with oodles of system memory. The idea that GPU renderers have to be limited to just the size of the video memory isn't really the case any more, as Redshift demonstrates. It also uses irradiance caching, photon mapping and other biased methods. So GPU renderers don't need to be stuck in slow-as-molasses naive path tracing mode, either.

Can that be because of Embree in Corona ? Everytime I switch to Embree in Vray, I get almost zero speedup in complex scenes (:-D I dunno) but my memory doubles instantly.
I've heard even Octane can be pretty efficient in what it allocates during rendering compared to CPU renderers but still. Regarding the 6GB, I've heard from one car-specific studio that the more complex car models only start at 12GB alone, so it would have to be
really simple product visualization.

But I'll have a look at this Redshift, it got me quite interested when it was posted here as comparison. Esp. the IC.

12GB car model? That sounds like a ton of unneeded stuff from importing the model, or extreme density settings when converting to polygons.. I don't know about cars, maybe Jeff Patton (who was here on the forum) would know.. But I remember I got from my employer designers some stuff that had a ton of crap in it. That was 90% not needed for rendering and made max almost unusable from import until i cleaned it up (and we're talking furniture here :D)... so basic optimization (deleting crap) made things usable for rendering with very small ram demands... So yes.. although still overkill, gpu rendering for some people still might look like a good option... especially if clients pay well enought to cover your investments and electricity bills :D

[Juraj]

I had quite complex models in mind, some with internals or fully detailed interiors. 12GB sounds quite right imho for me, not too far fetched.
Some quick examples:

[lacilaci]

I had quite complex models in mind, some with internals or fully detailed interiors. 12GB sounds quite right imho for me, not too far fetched.
Some quick examples:

Look, call me stupid or ignorant... :D but still doesn't look like 12Gb of polygons to me :D... I mean, it definitely wouldn't be 12GB of geometry when using subD modeling, sure a lot of geometry when using conversion from non polygon geometry... But I don't think stuff on those pics takes 12GB of ram.. Now maybe for super high resolution rendering, but I don't know... I actually would like to hear from someone qualified, someone that actually does handle those kind of models from actual designers... I simply cannot believe this wouldn't be possible with 6GB video card to render...  Geometry seems super smooth but still... 12GB of ram?... Again, unoptimized raw data loads of unused stuff, okay, maybe.... but if all you see on the rendering is all there is in the scene.. no.. no way.. I mean, I don't have experience with this kind of models... But i just somehow can't believe that :D.... sorry