my opinions about gpu use

[saudade]

I think that the CPU biased rendering method is "now" the only one that can reproduce the best results, noise free.

I stopped to buy gpus. I purchased new workstations instead.

Making clean and fast 3d renderings with only GPUs is almost impossible. Renderings have noise, there are heavy limitations using large textures, many procedural textures aren't supported, and many traps that we use to make biased renderings. No support for instances or proxies, particle, motion blur, plugins, etc.
I hope that Chaos will mix GPUs power with the CPU.

I watched a Cebas video that shows that mixing cpu+gpu power is possible to create biased noise free renderings.

[vlado]

I watched a Cebas video that shows that mixing cpu+gpu power is possible to create biased noise free renderings.

The performance boost that you can get in this way is very very limited (especially if you want to support the 3ds Max materials); at most 2x-3x, and sometimes less, depending on the scene. It would be much more effective to simply buy a new machine and use it for DR. This is because most of the time of a raytracer is not spent tracing rays, but calculating materials.

(You could argue that LuxRays achieves better acceleration; however this is because the shading is extremely simplified).

The best performance from GPUs can be achieved only by moving as much of the calculations as possible on the GPU itself, like we did for V-Ray RT.

Best regards,
Vlado

[renderfarmer]

VrayRT/GPU is still just an newborn yet is showing tremendous promise. I agree that it's not ready to take an advanced scene all the way to production, but it already serves a valuable role in our pipeline. We use it daily for lighting setups and whatever material previews RT/GPU is suited for at the moment given its limitations. That alone has saved us hours during the initial development phases of a scene. Considering the low cost of a GTX 590, and the high cost of an artist's time, the GPU's ROI was measured in hours... We still maintain a CPU renderfarm and don't foresee a near future without one. But we also see RT/GPU as a very cost effective compliment to it.

[saudade]

The performance boost that you can get in this way is very very limited (especially if you want to support the 3ds Max materials); at most 2x-3x,

2X-3X is not bad...

[AbnRanger]

On the same scenes with comparable settings/lighting, FR seems faster overall, to me. If you are looking at 20-30 min/frame with VRay and 10-15 with FR, then that 2-3x's is VERY significant. Now you are talking 5min renders instead of 20-30.

Having said that, finalRender R3-3.5 simply can't compete in the realm of interactive rendering. The IR crashes with such regularity and frequency it is practically useless, while RT is solid as a rock. I asked Edwin at Siggraph 2010 why the IR was so problematic and if they were going to do anything to fix it. He essentially pointed all the blame at Autodesk and ActiveShade. I think it's because his approach is more integrated (can use any material, including volumetric effects) and is therefore limited by Max itself...whereas RT operates more as a standalone, and isn't as reliant on Max to work properly.

So, there is the tradeoff. Stability=limited features. Full integration/features=instability.