GPU Render Slaves

[padre.ayuso]

So I'm seriously considering moving my production line to RT, however there is the subject of financing that is the main question. There are a few questions that will assist me in predicting my future with RT:

1. Per what I understand, RT does not linearly increase its speed with more GPUs. In other words, the difference between having 4 Titan x or 5 is minimal, other than you are spending $1,000 for minimal speed increase. Is this true? And is it subject to change in the near future to be able to get more speed when adding more GPUs?

Reasoning: A computer slave with 4 GPUs is only supported so far that I have found, by having 2 CPUs in the slave. The computer alone, without the GPUs is already costing about $5,000. Add the GPUs and we are looking at nearly $12,000. Is there a render node that can take 4 GPUs and only use one CPU?

2. Is RT able to DR to other CPU's in the NW? I understand the Hybrid Theory with CPU+GPU, can I use the GPU's in one machine, plus its own CPU, plus the CPU's in other machines? That may be a killer of an advantage.

3. What are the main goals in GPU technology from VRay in the near future? I want to be able to harness the speed of GPU, but at the same time Advance has been very stable so far for me and going RT may mean quite a long learning curve and adaptation curve.

Thank you for any data you may shed on this.

[benb32]

Hi there, I've browsed the forums a while and have seen the comment of diminishing returns crop up often. From what I understand Vray does scale in a linear fashion on both adv and RT. It's more useful to appreciate the math of the situation.

As an example, if you double the resolution of the picture you're rendering, you have increased the number of pixels of your image by 4 times. Double the number across and double going up - simply put think of sheets of paper, instead of a single sheet of a4 you need 4 sheets to double the width and height. So the render time should also go up by 4 times, what might take an hour will take 4 hours.

So if you double the number of CPUs (cores)/GPUs (Cuda cores) you will half your render time.

Going back to the example, a picture that initally takes 1 hour to render and 4 hours if your double the resolution. Let's say you had 2 GPUs (assuming both are equivalent). Each GPU was calculating half the picture, but doubling the resolution (thinking of the sheets of paper) now is calculating 2 sheets. 0.5 - 2 which is 4 times the work.

You are looking at the reciprocal when adding more, nodes or GPUs. To half the render time if you have 2 GPUs you need to double the number to 4. If you only have 1 GPU (1hour), 2 half the time (30mins), 3 third of the time (20mins), 4 quarter (15 mins). This is just the maths of the situation, nothing to do with Vray or indeed any other computationally intensive program.

Hope this makes things clearer

[padre.ayuso]

LOVL

Indeed I took math in school and got quite good at it and what you say makes sense, however, I have seen otherwise on the forum, from both users and developers. I believe Blago or Lele could shine some light on this.
Yet, VRay 4.0 with Optix Denoiser will make this part of the discussion totally irrelevant, I believe, as it seems that indee, with 3 or 4 GPUs I'll have the results I need in just but a few minutes. I hope this is true, and if so, I'll definitely move my production to GPU based, and make Advanced a thing of the past.
But I'd definitely like to test your theory out with RT and as soon as I have some spare time, I'll definitely do that. As a note, you are totally right on CPU being linear, the more cores you add, the faster it goes. But I don't believe this is true on GPU rendering with VRay RT.

[jblessing]

Hopefully you already got an answer for all of this on the fb group or else where. I see this question still asked a lot, so in our experience, the most cost effective route is to build 4 GPU/1CPU machines as slaves and render animations using backburner. DR for animations isn't as stable and the speed does not scale as linearly because of the overhead on each frame that the GPUs don't touch. The part of the rendering that GPUs do, does scale linearly (assuming you don't have any other bottlenecks in your PC). Check what GPUs are most effective (vray bench scores) before you just buy the most expensive. Right now, I would buy 4 x 1080ti's, or 4 x 1070's if I wanted to save money. This also depends on the size of your scene and if it will fit into the ram that is availble on those GPUs.

Then on each of your render slave boxes, you can have all of the GPUs and CPUs rendering in RT. In our experience, roughly 1 980ti = 1 5960x in speed. So you then could have 4 GPUs + 1 CPU rendering on each slave box, with an equivalent score of 5 GPUs. (Note that some scenes may be too demanding and won't render faster when using the CPU too.) Then when you want to scale up your rendering, just build more slave boxes.

I don't know if Chaos Group will go into additional specifics on the road map for GPUs, but previously they have pointed out the increase in performance for GPUs over the last few years is many multiple times better than CPU increases, so GPUs are the resource to utilize. Benchmarks of all kinds back this up, and with the demand for better GPUs for VR, this hardware trend won't stop any time soon.

[padre.ayuso]

Hopefully you already got an answer for all of this on the fb group or else where. I see this question still asked a lot, so in our experience, the most cost effective route is to build 4 GPU/1CPU machines as slaves and render animations using backburner. DR for animations isn't as stable and the speed does not scale as linearly because of the overhead on each frame that the GPUs don't touch. The part of the rendering that GPUs do, does scale linearly (assuming you don't have any other bottlenecks in your PC). Check what GPUs are most effective (vray bench scores) before you just buy the most expensive. Right now, I would buy 4 x 1080ti's, or 4 x 1070's if I wanted to save money. This also depends on the size of your scene and if it will fit into the ram that is availble on those GPUs.

Then on each of your render slave boxes, you can have all of the GPUs and CPUs rendering in RT. In our experience, roughly 1 980ti = 1 5960x in speed. So you then could have 4 GPUs + 1 CPU rendering on each slave box, with an equivalent score of 5 GPUs. (Note that some scenes may be too demanding and won't render faster when using the CPU too.) Then when you want to scale up your rendering, just build more slave boxes.

I don't know if Chaos Group will go into additional specifics on the road map for GPUs, but previously they have pointed out the increase in performance for GPUs over the last few years is many multiple times better than CPU increases, so GPUs are the resource to utilize. Benchmarks of all kinds back this up, and with the demand for better GPUs for VR, this hardware trend won't stop any time soon.

Thanks and good data to have. Since I started this post I saw the videos from Vlado on VRay 4.0 with OptiX testing, and it seems like moving at that point to GPU will be a wise move and I'm sure NVidia will have released their next super duper ti card or something of the sort.

Does anyone know where I can get a slave that will support 4 cards with one CPU per slave? Right now all I'm able to find are specs with dual 8 cores to be able to run 4 Cards the size of the Titan X.

Thank you!

[ramon_perez]

You can use a 7th Gen. i9 CPU which has enough PCIe Lanes (44) to run 4xGPUs. You can throw everything on a x299 board with 4 evenly spaced PCIE slots, like the Asus rampage extreme.

[werticus]

amd thread rippers are even better value when you look at how many pci-e lanes they have.