Rtx 4090

[glorybound]

This is how the usage looks in my testing, there is no difference between a 3090 and 4090 in GPU usage..

That isn't good news at all. A $500 card between a $1800 card and there is no difference in render times?

[francomanko]

He meant in GPU % usage, not rendertimes.

[Muhammed_Hamed]

That isn't good news at all. A $500 card between a $1800 card and there is no difference in render times?

GPU usage not render time, here are some render times for the RTX engine. Render time in seconds




Comparing the RTX 3080 to the RTX 4080 shows at least 1.8X improvement in render times. And Going from the 3090 to 4090 shows at least 2x improvement in render times
I tested the GPUs not connected to monitors, and all tests are run for at least 5 minutes(including scene compilation)

If you are getting less than 2X speed up with your 4090 compared to 3090, this is not expected to happen and you should start a ticket.
If you render time is 20 seconds, you will not get accurate results because scene compilation is taking longer than the actual rendering.

Best,
Muhammed

[glorybound]

Okay, that is better news. If they can figure out the usage, then it is all the better. I am good with 2X the speed

[Muhammed_Hamed]

Vray always uses 2x as much Vram on my 4090 than on my 3090.

I have checked 6 scenes today, on 2 of the machines at office. The memory usage is very close between the RTX 4090 and 3090 see the attachments below(the 3090 probably has more VRAM because it is connected to my monitors)
Do you have multiple Houdini instances open? If you monitor your VRAM usage does V-Ray flush V-Ray after you close your scene? sounds like a memory leak
Please start a support ticket, and mention the name of this topic in the title. We cannot reproduce this behavior on our side, so we will need to take a look at your setup via remote access if possible

And it's about 1.5x as fast on the 4090 than the 3090

This is not normal, you should get at least 2x faster renders with 4090. Otherwise, something is slowing down your scene and we need to take a look at your setup

The VFB keeps crashing to the point of rendering the program unusable at times - at least in Linux. This usually happens when displays are re-arranged in a multi display config. It's like it tries to place itself in a previous position which it can't and boom. And resetting the position will also cause a crash.This will then require the creation of a new render ROP in Houdini to get it stable again.

Sorry to hear you are having trouble with the frame buffer. The Houdini team chased this instability for a while around the time we released version 5.1 and we followed along with the users who reported this. To my knowledge there is currently no issues logged regarding the frame buffer positioning. Have you gotten in contact with us about this issue in the past?
We will need more information on this.

When my display isn't connected to gpu0 (the one in the 1st PCIe slot) and to gpu1 (3rd PCIe slot) instead Vray GPU won't render at all.

I cannot find anything about this in our system either, we need to reproduce on our side to be able to investigate. We have multiple machines with 4x GPU and 6x GPUs to try and reproduce
Have you tried to render with other engines in this case?
Please send us a ticket, if we reproduce on our end I will make sure it gets addressed, that is if it an issue with V-Ray GPU

Textures won't be rendered on GPU when motion blur is switched on, but only sometimes - like on Tuesdays because yesterday was Monday and everything was fine.

We deployed multiple fixes for this, it seems that there are still some cases where the issue happens like what you wrote in the other thread.
Please share a scene with us, If we reproduce on our side I will make sure it gets fixed

Why does it have to re-load the same assets for each frame in a sequence? A waste of time and resources and very annoying with larger textures.

This is on us, it is not an easy fix and requires research. We put effort into this in the past, and we will get back to this. It is still one of the most pressing requests
We are pushing on GPU Caustics now, once it is out we will get back to the topic of keeping resources between animation frames

Best,
Muhammed

[Muhammed_Hamed]

Vray benchmark even runs at 100% yet the render software doesnt

Hello,

Please check my comment about GPU usage, the benchmark is not consistent at 100% either. GPU usage will fluctuate on the CUDA test with high-end GPUs, which is normal
If you share a scene with me I will tell you what is going on, and if there is a way to get higher GPU usage on it

Taking an old scene with GPU lightcache is so problematic (infinite loop) an issue from many years ago

We have chased this instability for a while and solved many of these cases. Please start a ticket and send us the scene, if we reproduce on our side I will make sure it gets solved. Any crashes or hangs we give it priority in development

Moving around in IPR is so delayed its like it has to reload everything with every camera move.

Without GPU LC it should be instant, and the Densoiser could be used for even a better experience
Having long delays is not expected to happen and you should get in touch with us about it

Enmesh doesnt even work. Check old posts and states a year ago its in top priority list

It is currently WIP, whenever it is ready I will share with you.

IPR in vray has always crashed IMPO with itoosoft, Corona much more stable

I'm sorry to hear about your experience, the recent tickets we had over Railclone was solved by updating the plugin actually
I agree that making the experience stable is more important than adding tools or features. Please start a ticket and get us a scene, all crashes or hangs we treat with priority. If you need updates over a certain request or bug you submitted message me about it for updates

Thanks!

Best,
Muhammed

[leehumphries]

Muhammed_Hamed thank you for all the info. Massively appreciated
will check all your points and see how I get on
i did have loads of render elements so will check

[peter_hoehsl]

I have checked 6 scenes today, on 2 of the machines at office. The memory usage is very close between the RTX 4090 and 3090 see the attachments below(the 3090 probably has more VRAM because it is connected to my monitors)
Do you have multiple Houdini instances open? If you monitor your VRAM usage does V-Ray flush V-Ray after you close your scene? sounds like a memory leak
Please start a support ticket, and mention the name of this topic in the title. We cannot reproduce this behavior on our side, so we will need to take a look at your setup via remote access if possible



This is not normal, you should get at least 2x faster renders with 4090. Otherwise, something is slowing down your scene and we need to take a look at your setup



Sorry to hear you are having trouble with the frame buffer. The Houdini team chased this instability for a while around the time we released version 5.1 and we followed along with the users who reported this. To my knowledge there is currently no issues logged regarding the frame buffer positioning. Have you gotten in contact with us about this issue in the past?
We will need more information on this.



I cannot find anything about this in our system either, we need to reproduce on our side to be able to investigate. We have multiple machines with 4x GPU and 6x GPUs to try and reproduce
Have you tried to render with other engines in this case?
Please send us a ticket, if we reproduce on our end I will make sure it gets addressed, that is if it an issue with V-Ray GPU



We deployed multiple fixes for this, it seems that there are still some cases where the issue happens like what you wrote in the other thread.
Please share a scene with us, If we reproduce on our side I will make sure it gets fixed



This is on us, it is not an easy fix and requires research. We put effort into this in the past, and we will get back to this. It is still one of the most pressing requests
We are pushing on GPU Caustics now, once it is out we will get back to the topic of keeping resources between animation frames

Best,
Muhammed

I was taking the values for the Vram usage from the VFB which usually indicates nearly 2x for the 4090. I'm on Linux so there is no task manager and I was now looking at psensor instead which indeed shows a similar value for both cards which is inconsistent with the VFB but consistent with your task manager. Only one instance of Houdini but multiple render ROPs in my project.

With regards to speed this issue doesn't seem to be consistent. I was testing two scenes today, one with only a few VDB volumes and another with mostly refraction and a few textures on one object. Oddly the scene containing only the VDBs took 4m 25.2s on the 3090, 1m 16.6s on the 4090, and 1m 2.8s on both. This is in contrast to what was previously reported since in this case it's the 3090 that's a lot slower than expected. In the other scene however the times were 3m 15.2s on the 3090, 2m 25.3s on the 4090, and 1m 29.2s on both. Here I would have expected the 4090 to perform better.

I have a Houdini crash report for the VFB crash which I can send you (without the hip due to a very strict NDA). Let me know where I can upload it.

I will need to find some time to check the issue with the display hooked up to the 2nd GPU. It is a Threadripper system so both GPUs are in a 16x PCIe slot. This could be system specific, perhaps related to the mother board or BIOS and may not be reproducible on another system. It's just that with other renderers there doesn't seem to be an issue.

The corrupt texture issue with motion blur on GPU is not consistent. It happens in one project and not in another even with similar configurations. I had this happen a few weeks ago, then it went away, and on Tuesday it popped up again - all on an asset that hadn't changed in all that time. When it does happen however I discovered that it seems to be related to the presence of a Cd attribute on the geometry regardless of whether that's used in the shader or not. When I delete the Cd attribute the issue is gone and the render turns out fine. I've had issues in the past when using Cd in the diffuse channel in a shader where with motion blur the color was ignored and the object rendered black. I can't share this project at this stage.

Sharing resources between frames would be really cool for scenes with large textures. Look forward to that feature.

Thanks, Peter


EDIT: Earlier today I was rendering out a sequence which I stopped half way to do other things. When I restarted the render for the 2nd half the textures were messed up again. They were fine earlier and nothing has changed in the scene. I found and deleted a Cd attribute and things are fine now. Perhaps Vray is not just extracting only the attributes it needs but accidentally loads data that's not used which ends up as a copy of the geo in RAM?

[glorybound]

I am testing my 4090, and it is slow. CPU is rendering my scene in less than 10 minutes and the GPU, 35 minutes. How many GPU buckets should I be seeing with the 4090 RTX? Also, since I see DR in the GPU render settings, I am assuming you can DR using GPU, is this correct?

[s_gru]

someone correct me if im wrong, but i remember reading that bucket rendering was recommended in vray v5 but when i noticed it took forever i asked about it and someone at choas said use progressive. something had changed in newer release of v5... or it was old documentation? something. but anyway i have always had fastest gpu rendering on progressive.

edit yes you can use DR but need a node license on the other machine

[glorybound]

it is the same node as any other node, correct? So, with DR GPU, use Progressive?

[Muhammed_Hamed]

We recommend using Bucket mode for DR and Progressive mode for everything else, Progressive mode uses less GPU memory and could be faster in some cases. But for DR, it uses a lot more network bandwidth than bucket mode

glorybound Is it possible to take a look at your scene, under NDA
muhammed.hamed@chaos.com

Thanks!

[s_gru]

ah yep i remember the specifics now. muhammed i think you were the one who answered me back then.

i no longer have a rnode so dont even use dr anymore.

We recommend using Bucket mode for DR and Progressive mode for everything else, Progressive mode uses less GPU memory and could be faster in some cases. But for DR, it uses a lot more network bandwidth than bucket mode

glorybound Is it possible to take a look at your scene, under NDA
muhammed.hamed@chaos.com

Thanks!

[glorybound]

I ran the nutcraker scene and GPU rendered in 5 minutes. With CPU DR it rendered in 12 minutes.

[dlparisi]

I ran the nutcraker scene and GPU rendered in 5 minutes.

which scene is this? Link? I'm interested in comparing.

EDIT: Nevermind, I saw it's from the other thread: https://forums.chaos.com/forum/v-ray...83#post1190883