Muhammed,
as your gpu shaderball testscene renders fine but our scenes don´t we need some assistance in converting our existing scenes and workflows to gpu rendering. We we now save your shaderball rendersettings and try them with our scenes..
Matthias
Muhammad? What to do now?
Sorry, I just saw your post. I will move this thread to the correct forum now
Swap GI to BF and use Cuda mode in performance tab. Let me know if this helps
Could you please share more information on your scene setup? Does it use proxies, displacement, FP?
Best,
Muhammed
Muhammed, sorry, didnt see the vray gpu forum (quite at way scrolling down the screen
I found the vray gpu list: Chaos Docs
does that mean that if we use any of the incompatible features the scene wont render? (Using vraymap for showing blueprints in viewport but they are of course hidden while rendering)
Hi @digitarch
V-Ray GPU is a separate renderer, optimized for speed and built from scratch to utilize the GPU hardware. There are known limitations in the link you shared, it is best to start your new scene using the GPU engine from the start. This way you avoid any differences or limitations on swapping
It is also possible to render your current scenes with V-Ray GPU, with some tweaks. Did you swap to Cuda and try with your scene?
Some of the errors you got above are because of exceeding the GPU memory. Is it possible to upload me one of your scenes? I will tell you what is causing the errors and how to approach this
Best,
Muhammed
Muhammed, Swap GI to BF and use Cuda mode in performance tab didn´t help. We start now with a new project from blank screen for testing gpu performance and see …
Good luck!
Please let me know if you have questions on the GPU workflow
Best,
Muhammed
Hello Muhammed,
thank you for your help, i really apprechiate it.
I now tried to compare GPU and CPU rendering in my current project by using VRayMaterial Override with a grey Material to keep things simple.
I did a lot of test rendering with different Noise Thresholds comparing the rendertimes.
Here are some results:
V-Ray GPU 5 Bucket Mode, CUDA, Brute Force, Samples limit 3000, Noise limit 0,2, RTX only
1min16sec
V-Ray GPU 5 Bucket Mode, CUDA, Brute Force, Samples limit 3000, Noise limit 0,2, hybrid
1min6sec
V-Ray 5 Bucket Mode, Brute Force, Max Subdivs 100, Noise threshold 0,2
43 sec
V-Ray GPU 5 Bucket Mode, CUDA, Brute Force, Samples limit 3000, Noise limit 0,05, RTX only
5min00sec
V-Ray GPU 5 Bucket Mode, CUDA, Brute Force, Samples limit 3000, Noise limit 0,05, hybrid
3min55sec
V-Ray 5 Bucket Mode, Brute Force, Max Subdivs 100, Noise threshold 0,05
3min16sec
V-Ray GPU 5 Progressive, CUDA, Brute Force, Samples limit 3000, Noise limit 0,2, RTX only
57sec
V-Ray GPU 5 Progressive, CUDA, Brute Force, Samples limit 3000, Noise limit 0,2, hybrid
1min11sec
V-Ray 5 Progressive, Brute Force, Max Subdivs 100, Noise threshold 0,2
57sec
V-Ray GPU 5 Progressive, CUDA, Brute Force, Samples limit 3000, Noise limit 0,1, RTX only
2min5sec
V-Ray GPU 5 Progressive, CUDA, Brute Force, Samples limit 3000, Noise limit 0,1, hybrid
1min35sec
V-Ray 5 Progressive, Brute Force, Max Subdivs 100, Noise threshold 0,1
1min38sec
As you can see in Bucket Mode CPU rendered much faster than GPU in both cases did (RTX as well as hybrid)
Using Progressive GPU Hybrid rendered a bit faster than CPU, but RTX alone still remained very slow.
Now our question is whether we didn’t set up GPU rendering correctly?
I attached a Link of the scene and both cpu and gpu rendersettings for you to download:
Thank you in advance.
Best regards, Marco
Thank you for the scene, @digitarch
Here are a few tips to help with your testing,
-Your test runs need to be at least 5 minutes for accurate results, V-Ray GPU takes a few second to move geometry and texture to system memory then GPU memory. So if your render time is 1 minute, half of that is probably the process of moving the scene into VRAM.
-Use 3000 LC Subdivs for interior scenes, this affects both V-Ray and V-Ray GPU, specially the GPU engine, this can double your rendering speed.
-Use the sky texture in a domelight not 3Ds Max environment slot, the Domelight is always faster (Applies to both V-Ray and V-Ray GPU)
-Disable displacement from Global Swatches, unless you manually match the look between the 2 different engines. Right now displacement looks so different between the 2 engines.
-Delete render elements for such tests to be safe
-Use RTX mode, it is super fast for the type of scene. The hardware acceleration is very helpful
-Your goal should be achieving the same noise results, the GPU engine gives cleaner results and slightly brighter by default. You need to keep that in mind, you need to adjust your noise threshold to make for a fair comparison. Check below
I used slightly higher resolution in my testing, to make sure the testing is accurate
Clay test,
V-Ray Bucket mode, 5950X CPU 16 cores/32 threads >> 49 Minutes 31 Seconds
V-Ray GPU Progressive mode, RTX 3090 >> 3 Minutes 17 Seconds
For this test I used .01 noise threshold for V-Ray and .02 noise threshold for V-Ray GPU, and still the result of V-Ray GPU is cleaner
Beauty test,
V-Ray Bucket mode, 5950X CPU 16 cores/32 threads >> 58 Minutes 43 Seconds
V-Ray GPU Progressive mode, RTX 3090 >> 10 Minutes 4 Seconds
.01 noise threshold for V-Ray and .02 noise threshold for V-Ray GPU
So why am I comparing V-Ray’s Bucket mode to V-Ray GPU progressive mode?
This is the recommended way of using each renderer, V-Ray(CPU) is fastest and most efficient with bucket mode and V-Ray GPU is fastest with progressive mode
Have a great weekend!
Best,
Muhammed
Hello Muhammed,
Thank you so much for the detailed response!
We will try your settings and see if we can reproduce it.
Best regards, Marco
Good morning @digitarch
Let me know how this goes, your 3060 should be exactly half of the 3090’s performance. You can match your results with mine above, it should be close
Best,
Muhammed
Hi, found this thread really important for us,as our study is thinking to move on GPU rendering too.
Not that much info about GPU rendering, on this forum and even on the internet, at least not very delighting one.
We really have several FAQS about this issue, and we didnt achieve to find out simple answers, although we think they have to be somewhere.
First of all, sorry Muhammed, but we find out not a fair comparision the one you are just showing up.
· RTX 3090 cost around 2.500 euro, but 5950X just 500 euro. A better comparision should be RTX 3090 with Amd Ryzen 3970X.
· We can accept the point of view with the noise thresold, but both has to been render with the same method, progressive method consumes more memory and has some limitations, so it is also an unfair comparision.
This is more or less what we have find out on the internet. Not fair comparisions, and not clear info about some important points that we gonna point out here, so perhaps Muhammed and chaosgroup can put some light on it..
The first point would be to compare with the same method (progressive or bucket) and with two equal price piece of hardware, GPU and CPU. This can give us a good starting point to know, where ever we have to spend more money pon GPU or CPU to get speed improve. We havent find nothing close to this on the internet. And it all about money. Where we should put all our budget in.
Second point, is assuming GPU will be faster than CPU (we dont really know cause we cannot get a decent test with equal hardware and equal method), we get to an important limitation about VRAM memory on GPU.
Right now, a normal scene on our study consumes about 100 - 140 GB RAM. Lots of polygons, high resolution textures, displacement and high resolution final image.
No graphic card has this amount of VRAM, we know it seems that GPU needs less VRAM than CPU does, but do we have a good criteria to know if we can render our scenes with GPU?, or do we have to spend thousands of euros on a workstation with the latest Graphic cards, and after that hit render and get memory exception error?
Some of our questions are regarding this matter. And it would be greatful that someone in Chaosgroup put some info on this.
We have read a lot about this matter, and we have come close to confirm that some Nvidia cards with NVlink can share their VRAM memory. It seems, but havent get an official reply, that Vray 5 is ready to handle it. Is it so?
How much Vram do we need, when we need 120 GB for CPU? More or less, nothing like, it depends on bla,bla,bla,… . Just a rough value. So we come close to know how many graphic cards do we have to link with Nvlink.
Next, if VRam is not enough, Vray is compiled to work with the Ram memory of the computer, even if it slows a bit the render, or you just cannot render the image? Cause i really know that if i have n AMD Ryzen 3990X with 256 GB RAM i can render whatever i want, pretty sure aobut this. It can take a night long, but i can render it on high resolution image. Is it the same with GPU? Or we get a point where we have spend 8.000 euro on a workstation and we hit render and get an error and no final render?
This are some of our questions that need to get an answer before we decide to move on GPU Rendering if finally we do so.
Thank you.
Hey @Witt3D and thanks for the feedback, I will comment on some of the points below
We are in the process of updating our GPU documentation, there will be a section about this topic soon.
I tested on a Threadripper 3970X as you requested, here are the results
Clay test,
V-Ray Bucket mode, 3970X CPU 32 cores/16 threads >> 28 Minutes 2 Seconds
V-Ray GPU Progressive mode, RTX 3090 >> 3 Minutes 17 Seconds
Beauty test,
V-Ray Bucket mode, 3970X CPU 32 cores/16 threads >> 33 Minutes 48 Seconds
V-Ray GPU Progressive mode, RTX 3090 >> 10 Minutes 4 Seconds
So, the same conclusions from above are correct. a 3090 is much faster than a Threadripper 3970X or even the 3990X which is +4000 Euros
One note is that a 3090 is 1800 Euros here, and prices are going down everywhere. In some places it is even closer to MSRP
A 5950X is 850 Euros here, I have never seen it near 500 Euros..
On the other hand a 3970X is out of stock in most places now. You will only find the 3975WX Pro which is similar in performance but more expensive, it is +3000 Euros here
A second note is that a Threadripper requires quad channel memory, expensive cooling and expensive board, this all adds up to the cost
A third note, you can only house a single Threadripper CPU per machine, while it is possible to use multiple RTX GPUs in the same system. Up to 4 per machine without issues, and expandable to 7 with Riser cables. This expandability is one big pro on the GPU side
Anyways, I used this CPU in my testing above because it is a consumer board which is accessible to most people.. it doesn’t require an expensive board or cooling.
This is not correct, progressive mode is very reliable on GPU and it is the recommended way of working. It is faster than Bucket mode in most cases, and uses less GPU memory.
Keep in mind that V-Ray and V-Ray GPU are different render engines, they produce different noise results. The GPU engine was written from scratch to be as fast as possible and to unleash the GPU hardware.
The GPU engine will have a brighter and cleaner result, you can get away using a higher noise threshold in many cases.
I worked as a concept designer/visualizer in Print for more than 6 years. I transitioned my previous 2 studios Renntech and Carlex into GPU rendering fully, and both are still stuck to GPU rendering until now. I did a lot of research on this topic, and eventually convinced the team leads into making the transition. I have production experience with many renderers including V-Ray, Octane, Redshift, Arnold and others, I have seen it all 
I joined Chaos last September as a GPU product specialist, we do performance testing regularly between V-Ray and V-Ray GPU.
System memory and GPU memory don’t work in the same way. Some of my scenes that take 70-80 GB of system memory would render fine on 11 GB GPUs
V-Ray GPU is quite efficient on using GPU memory, if you send me a sample scene I can test and show you the memory footprint between CPU and GPU. I have dual 3090s NVlinked, I think 48 GB of GPU memory are going to work in your case.
Another note is that V-Ray GPU and other GPU renderers will use a lot of system memory, this is normal. You ideally need 128 GB of system memory for a dual-3090 setup
Yes V-Ray GPU supports NVlink and memory pooling, this applies to version 5 and version 4
At least dual 3090s + NVlink to be safe
You could send me a sample scene and I will show you the numbers
This is not supported yet, it will come in a future update. We call it OOC Geometry/Textures, where V-Ray GPU will use system memory when it runs out of GPU memory
a 3990X + 256 GB of memory will cost you 7000 Euros at least, it is not a cheap setup.
On another note, V-Ray GPU is quite reliable now, you will be safe as long as you start your projects from scratch using the GPU engine. It is being used in production in many studios, Porsche, Dabarti and others
Best,
Muhammed
I tested on a Threadripper 3970X as you requested, here are the results
Clay test,
V-Ray Bucket mode, 3970X CPU 32 cores/16 threads >> 28 Minutes 2 Seconds
V-Ray GPU Progressive mode, RTX 3090 >> 3 Minutes 17 SecondsBeauty test,
V-Ray Bucket mode, 3970X CPU 32 cores/16 threads >> 33 Minutes 48 Seconds
V-Ray GPU Progressive mode, RTX 3090 >> 10 Minutes 4 Seconds
Hi Muhammed, thank you for the new test. Good to know this values.
This expandability is one big pro on the GPU side
We have read somewhere that some graphic cards can just be link on pairs with Nvlink, it means that you just can put 2 together. I dont know if that is so with 3090 or not, just can you confirm that you can put more than 2 3090 together? Thank you.
This is not correct, progressive mode is very reliable on GPU and it is the recommended way of working. It is faster than Bucket mode in most cases, and uses less GPU memory.
Keep in mind that V-Ray and V-Ray GPU are different render engines, they produce different noise results. The GPU engine was written from scratch to be as fast as possible and to unleash the GPU hardware.
The GPU engine will have a brighter and cleaner result, you can get away using a higher noise threshold in many cases.
Ok, still for us an unfair comparasion. Not same rendersettings, not same render engine, you told is rewritten from scratch, not same limitations on each render engine, so for us is simply not a fair comparasion. True ones have to be done under exact same conditions and with the exact same result, not slightly different, but anyway we cannot see any on the internet cause we guess that GPU and CUDA under Bucket mode will be slower than CPU.
I worked as a concept designer/visualizer in Print for more than 6 years. I transitioned my previous 2 studios Renntech and Carlex into GPU rendering fully, and both are still stuck to GPU rendering until now. I did a lot of research on this topic, and eventually convinced the team leads into making the transition. I have production experience with many renderers including V-Ray, Octane, Redshift, Arnold and others, I have seen it all
I joined Chaos last September as a GPU product specialist, we do performance testing regularly between V-Ray and V-Ray GPU.
Thank you, we appreciate your experince and means a lot for us to take the right decision.
System memory and GPU memory don’t work in the same way. Some of my scenes that take 70-80 GB of system memory would render fine on 11 GB GPUs
V-Ray GPU is quite efficient on using GPU memory, if you send me a sample scene I can test and show you the memory footprint between CPU and GPU. I have dual 3090s NVlinked, I think 48 GB of GPU memory are going to work in your case.
GReat to hear that!!!
At least dual 3090s + NVlink to be safe
You could send me a sample scene and I will show you the numbers
WOW, that would be awsome!! Should be send a vrscene file, the point is that some textures would not be accessible, cause inside our scene there are lots of small vrscenes files with their own textures paths.
This is not supported yet, it will come in a future update. We call it OOC Geometry/Textures, where V-Ray GPU will use system memory when it runs out of GPU memory
Again that is superb!!! REally good news on that, and should make it easier to change to GPU.
Thank you again for all the info. So it will be easier to take the right decision.
Best.
And sorry, you just tell us where to buy to get your list of prices. Would be great. Here is what we get:
Ryzen 5950 for 556 euros.
https://www.google.es/shopping/product/1241695243945151256?q=5950X+ryzen&client=safari&hl =es-es&biw=414&bih=719&tbs=vw:l&prmd=snvi&prds=eto:377 8454643158039002_0,cdl:1,prmr:1,rsk:PC_13904683267 290411097,cs:1
RTX 3090 nothing less than 2100 euro
[Amazon.es : rtx 3090
[/URL]a 3990X + 256 GB of memory will cost you 7000 Euros at least, it is not a cheap setup.[[/QUOTE]](Amazon.es : rtx 3090)
Seriously, we have two different budgets here for two different workstations with 2 rtx 3090. Any of them is less than 7.000 euros. And again, for production i can guarantee with the above setup you can render anything you want. With GPU system right now you cannot confirm that.As mentioned before, for us would be lovely, great, awsome, incredible to move on a system where we have to spend less money and get much faster render times. WOW, this is magical and our dream.
The point is to know in a certain way that you can move your porduction to GPU without any great problem. As told you before, should be really painful and inadmissible to spend 6.000, 5.000, 7.000 euros on a GPU Workstation and after that be impossible to render some scenes. With the other config that wont happen at all.](Amazon.es : rtx 3090)
Hi again,
after another round of testing, everything works fine with simple scenes. But if I try to use GPU on a normal work project it crashes, does not render at all or renders black.
Cleaning up the scene helps a bit, but I still cannot render a complex scene with GPU. Especially Itoo Forest objects and Vray proxies (foliage) seem to bring GPU to its limits.
What is the best method to set up a more complex scene with GPU rendering?
We are still testing with the GeForce RTX 3060 with 12 GB
Best regards.
***********