Farm upgrade questions about Vray RT/Next

[ben_hamburg]

Hi there,

we?re about to upgrade our small render farm and i was tasked looking into the current state of GPU rendering, to determine wether we should invest the extra bucks and good graphics cards for the new render nodes.

I already checked the VrayRT supported features list and it seems like a lot has been done here since I last checked.

So here are my questions:

1. Biggest problem so far was memory as far as I remember. If the scene was too big, it didn?t fit in the GPUs memory and could not be rendered, which was the case for most of our scenes back then.
Is this still the case or can this be circumvented by using the new hybrid rendering method?
2. If a feature is not supported by Vray RT GPU, but is supported for Vray RT CPU - will it still render, when both are enabled?
3. I didn?t see the Vraydisplacement modifier - is it supported or not?
4. From a general experience: Is it worth it? In a bang for bucks calculation - should we just invest in more render nodes or in combined CPU/GPU render nodes? What are your experiences using it in production? Are there many hickups to be solved or is it generally shoot and forget, like it is with Vray otherwise?

All comments appreciated!

[super gnu]

well im not a Dev, but having switched mostly to GPU recently, i can give my insight.

1. memory is still a major issue.. with no sign of reasonably priced, decent gpus with more than 12GB even on the horizon (id love to be corrected) i do still often find jobs which i cannot render on GPU.

having said that, its tons better than it was, textures can now be streamed from Ram/Disk if they are too big to fit in GPU ram. so the main culprit nowadays is too much geometry. im often surprised when a scene im sure will fall over actually manages to render.


-the new hybrid rendering can be used, but you would be limited to cpu only, sicne the gpu would still be unable to contribute if the scene is too big. hybrid mode basically adds a CUDA driver for the cpu, so treats it as another gpu.

-if you are limited to cpu only i think normal vray would be faster.


2. if a feature is not supported by RT gpu (i.e. CUDA) then it also wont work on the cpu with CUDA driver. in that case swap over to RT cpu or better, normal vray.

-ive found that most of my daily used features now seem to be covered by RT, and i can even open older scenes with a reasonable chance they will render ok.. (edit: ive got into the habit of avoiding procedural textures, some work, some dont, the baking that rt gpu does on incompatible textures to render them is usually not high enough resolution.. id like a control for this in vray..)

3: displacement is supported, but it tends to eat gpu ram quite heavily, as (unless its changed recently) it generates a pre-tesselated mesh for the whole displaced object on cpu , then uploads it to GPU, as you can imagine, that can get quite heavy.

4. from my point of view yes. definitely worth it. but ive got quite a lowly workstation (6 core 4.6ghz) and a single titan X (second arriving next week the titan X often renders several times faster than my workstation, so its a no-brainer for me. when im forced to go back to cpu rendering i either do overnight renders or farm it out.

-on that note, ive found that renderfarms offering GPU services are quite expensive. hard to be sure but id say more expensive per-frame than normal cpu rendering despite the better bang for buck of the hardware.

[savage309]

Good answers, thats super_gnu
I would like to give my view as well:
1. V-Ray GPU is very efficient on memory. We use very little for textures. Geometry is still a problem that we look at. If the scene can't fit on the GPU, you can still render it with the Hybrid approach, using the CPU as a CUDA device. We guarantee that the result will match completely. It is like an insurance that even if the GPU can't handle a scene, you still can get the job done (although a bit slower).
2. We have two engines now - V-Ray and V-Ray GPU. V-Ray runs on CPUs only. V-Ray GPU runs on GPUs and CPUs (emulates them efficiently as CUDA devices). If the features is not in V-Ray GPU, you will have to manually switch to V-Ray. We don't try to make V-Ray and V-Ray GPU render 1:1 the same. They often render close, but always will be different.
3. 3D displacement is supported. The displacement algorithm in V-Ray GPU 3.6 and older really some memory to work. This is drastically changed in V-Ray GPU Next and it is very memory efficient now.
4. As usual - we suggest you to try it yourself before investing a lot in hardware. You can use V-Ray GPU even without CUDA device - the Hybrid approach emulates the CPU as a CUDA device. It will be slower, but the results that you see in the render will match completely the GPU results. Here is how to enable it in V-Ray 3.6 https://www.youtube.com/watch?v=rf1MkuflH3M.

Keep in mind that V-Ray GPU works great with GTX GPUs as well (https://benchmark.chaosgroup.com/gpu).

Best,
Blago.

[ben_hamburg]

Thanks for the input...based on that I think we´ll go with CPU for now and start testing Vray GPU on productions first at a small scale. We can still upgrade them with big GPUs and gain a big boost, if it works for us.
A big segment of our work is medical animations, for which both displacement and procedural textures play a big role, so I still have my concerns about these ...