Bad FPS with 135 mln polys

[charlyxxx]

Hi, dear support =)

So =) I hope you remember this post

I tried again to open my scene in Beta 2 and the first 5 minutes my Desctop was frozen and I couldn't do anything, and after that, it was opened!!! Yeah!

But, I have a very bad FPS on my 2x2080TI..

Video with issue -> https://www.youtube.com/watch?v=2W4b...ature=youtu.be

As you remember, in 3ds max everything is ok -> https://www.youtube.com/watch?v=H2Cj...ature=youtu.be

Scene -> http://bit.ly/2P1jQuY

Thank you and have a great day.

Best regards,
Andrew.

[vladimir.nedev]

It looks like you ran our of GPU memory.
Windows 10 and DirectX 12 seem to have a virtual memory system, which swaps data in and out of the GPU automatically.
But, of course, this is very slow, so that's why you get the 5 minute freeze and the very low FPS after that.

I mentioned it in the other thread, Lavina needs more memory for a particular mesh compared to 3ds Max, because with ray-tracing we need to keep the ray tracing acceleration structure in GPU memory, in addition to the vertex positions/normals/UVs.
It's also possible that 3ds Max streams the mesh in, something that's not really feasible with a ray-tracer.

We have an issue logged to improve performance with high triangle count meshes, but it will take some time.
And even after this improvement, it's very much possible that your mesh won't fit in GPU memory.

In this case, we have to implement another improvement, where Lavina would evict "less important" geometry and textures from GPU memory.
The evicted meshes would be rendered as bounding boxes, and the "importance" would be determined based on whether you see the mesh directly or as part of GI calculations, etc.
This, of course, means that part of the image won't have final quality.

Some users have mentioned an "out-of-core" solution, like the V-Ray Proxies.
This is not feasible for a real-time solution like Lavina, where we have to go through and render the whole frame with at least 20 FPS.
"Out-of-core" works with offline renderers because they can concentrate on small parts of the image (buckets) for a longer period of time, which requires less memory.

For now, the only solution for your scene is to have a GPU with more memory.

Greetings,
Vladimir Nedev

[charlyxxx]

Hi vladimir.nedev

It looks like you ran out of GPU memory.
Windows 10 and DirectX 12 seem to have a virtual memory system, which swaps data in and out of the GPU automatically.
But, of course, this is very slow, so that's why you get the 5-minute freeze and the very low FPS after that.

Hm, but I have 2x2080Ti with NVlink and saw, that Lavina used 13Gb in this project. Usually, when I have out of ram - I have a crash, like as in the previous beta... But as I understand, you added NVLink support. Right? That is I see 13 GB using from my 22 GB.

I mentioned it in the other thread, Lavina needs more memory for a particular mesh compared to 3ds Max, because with ray-tracing we need to keep the ray tracing acceleration structure in GPU memory, in addition to the vertex positions/normals/UVs.
It's also possible that 3ds Max streams the mesh in, something that's not really feasible with a ray-tracer.

Yes, I remember this =) I just try to understand, for what I can use the Lavina =) Because many subscribers asked about this too.

We have an issue logged to improve performance with high triangle count meshes, but it will take some time.
And even after this improvement, it's very much possible that your mesh won't fit in GPU memory.

I hope this will be possible with 2x2080TI + NVlink =)

In this case, we have to implement another improvement, where Lavina would evict "less important" geometry and textures from GPU memory.
The evicted meshes would be rendered as bounding boxes, and the "importance" would be determined based on whether you see the mesh directly or as part of GI calculations, etc.
This, of course, means that part of the image won't have final quality.

Did you mean LODs or Distance-Based Tesselation? Like as we have in Unreal Engine. When the user saw a beautiful tesselation and details in front (near) and bad lowpoly details in background ( far)

Some users have mentioned an "out-of-core" solution, like the V-Ray Proxies.
This is not feasible for a real-time solution like Lavina, where we have to go through and render the whole frame with at least 20 FPS.
"Out-of-core" works with offline renderers because they can concentrate on small parts of the image (buckets) for a longer period of time, which requires less memory.

Ok. I understand this, thank you for this detail info and explanation.

For now, the only solution for your scene is to have a GPU with more memory.

So, as I understand. If I want to work with large landscapes, like this - I should sell my 2x2080Ti's and bought 2-4 Nvidia Quadro RTX? =) Hm, this is just a test =) I think we don't need 135 mln polys in scenes like this, because we will have V-Ray Displacement in the next versions. Right?

Thank you and have a great day.

Best regards.
Andrew.

[vladimir.nedev]

Hm, but I have 2x2080Ti with NVlink and saw, that Lavina used 13Gb in this project. Usually, when I have out of ram - I have a crash, like as in the previous beta... But as I understand, you added NVLink support. Right? That is I see 13 GB using from my 22 GB.

No, we don't support NVLink, and I don't think it's even possible through the DirectX raytrcing API.
This might be a feature that will be available only with V-Ray GPU where we use CUDA/OptiX.

It's a bit random whether you get an out of memory crash, or the rendering is just slow and the data is streamed in and out of CPU memory.
We are not sure yet why this happens.

Did you mean LODs or Distance-Based Tesselation? Like as we have in Unreal Engine. When the user saw a beautiful tesselation and details in front (near) and bad lowpoly details in background ( far)

Yes, something similar.

Greetings,
Vladimir Nedev