OctaneRender™ 2020.1 RC1 [superseded by 2020.1 RC2]

[nejck]

Is there a difference in running NVLink on consumer cards (Geforce) versus Quadros in Octane? Afaik they support NVLink with TCC mode on but I'm not sure how the implementation in Octane works.

As per the latest NVidia drivers while writing these lines the only difference is that multiple NVlink setups are just supported for Quadro cards. This seemed to work in the past for some users but NVidia confirmed it is not supported and will not be in the foreseeable future.

Thank you Mojave, your answer is much appreciated

[Dmikucki]

The Random Walk medium still shows seams when two materials are used on the same object. This is a regression from the old node. It seems to happen even with no textures and a duplicate material node. See screenshots.

Does this happen by any chance just when you enable RTX acceleration support in the device settings? We have internally solved a few issues that didn't make it into this release that could be causing this. Would you be able to share this scene or a trimmed down version we can use to test your report? Thank you.

I don't have an RTX card, so I can't answer the former question. As for a test scene, here is a very simple version of one.

[mojave]

Oops, sorry. I tried several times to upload the orbx file, but the forum keeps cutting it off at about the halfway point. The file is 492 MBytes in size.
Jason

Thank you for sharing your ORBX via PM.

I can confirm there's a performance regression between 4.05 and 2019.1/2020.1 for this project we will be looking into.

Also your scene has denoiser enabled with an interval of 20s, if you are looking for best performance you can use "Denoise on completion".

[grimm]

Oops, sorry. I tried several times to upload the orbx file, but the forum keeps cutting it off at about the halfway point. The file is 492 MBytes in size.
Jason

Thank you for sharing your ORBX via PM.

I can confirm there's a performance regression between 4.05 and 2019.1/2020.1 for this project we will be looking into.

Also your scene has denoiser enabled with an interval of 20s, if you are looking for best performance you can use "Denoise on completion".

Thanks Mojave,

Jason

[Notiusweb]

Anime Kernel and Polarized Lighting -
Did they cause too many coding conflicts and are now abandoned?

[Goldorak]

Anime Kernel and Polarized Lighting -
Did they cause too many coding conflicts and are now abandoned?

No, they are still being actively worked on, but we prioritized more urgent user requested features for 2020.1 based on feedback, like ACES, C4D GPU noises, spectron quad/sphere primitives, random walk SSS and optix 7 RTX backend. Anime and brigade kernels are also being developed on RNDR SDK which is still waiting on stable Vulkan and Metal drivers.

[karl]

-The settings for the workstation :
M/B:supermicro x10DRG-Q, CPU: 2x E5-2680v4, RAM:128G RECC , GPU: 4x Asus 2080ti turbo version, NVlink bridge: 2slot for Nvidia quadro 6000, only 2 2080ti is bridged

-Octane Settings:
RTX: off, out of core: off, network rendering: off, GPU head room 800MB(1024MB is the maxmum for my hardware settings)
during the process I may changed these switches on and off, and finally the image I post appears...

-The scene
the scene as displayed contains 44Million triangles: there are trees and a piece of carpet, which contains half of the triangular counts,
the orbx file is about 4GB, obviously too big to be uploaded, and may not worthy, since the carpet is made of 3ds max's hair and fur, the 3dsmax file is small enough and good for transmitting, in 3dsmax to adjust the parameter "hair count" you can get as many triangles as you want (a disaster for project yet good for testing)
so Iam uploading the carpet max file instead of the orbix
(BTW, some version for 3dsmax plugin will fail when convert the fur, need to use a right version...)

triangluar_generator.rar

Jim

Hi Jim,

Thanks for uploading the scene. The reason why this scene is failing to render even with NVLink enabled is because it's a single large mesh rather than multiple smaller meshes. Currently, memory pooling only allows us to allocate a mesh on one GPU and share the allocation with another GPU, but each mesh still needs to fit in its entirety on at least one GPU. We plan to look into automatically splitting meshes in the future but for now the only solution is to somehow reduce the mesh size, either by manually splitting into multiple smaller meshes or reducing the overall triangle count.

[jimho]

-The settings for the workstation :
M/B:supermicro x10DRG-Q, CPU: 2x E5-2680v4, RAM:128G RECC , GPU: 4x Asus 2080ti turbo version, NVlink bridge: 2slot for Nvidia quadro 6000, only 2 2080ti is bridged

-Octane Settings:
RTX: off, out of core: off, network rendering: off, GPU head room 800MB(1024MB is the maxmum for my hardware settings)
during the process I may changed these switches on and off, and finally the image I post appears...

-The scene
the scene as displayed contains 44Million triangles: there are trees and a piece of carpet, which contains half of the triangular counts,
the orbx file is about 4GB, obviously too big to be uploaded, and may not worthy, since the carpet is made of 3ds max's hair and fur, the 3dsmax file is small enough and good for transmitting, in 3dsmax to adjust the parameter "hair count" you can get as many triangles as you want (a disaster for project yet good for testing)
so Iam uploading the carpet max file instead of the orbix
(BTW, some version for 3dsmax plugin will fail when convert the fur, need to use a right version...)

triangluar_generator.rar

Jim

Hi Jim,

Thanks for uploading the scene. The reason why this scene is failing to render even with NVLink enabled is because it's a single large mesh rather than multiple smaller meshes. Currently, memory pooling only allows us to allocate a mesh on one GPU and share the allocation with another GPU, but each mesh still needs to fit in its entirety on at least one GPU. We plan to look into automatically splitting meshes in the future but for now the only solution is to somehow reduce the mesh size, either by manually splitting into multiple smaller meshes or reducing the overall triangle count.

Hi Karu
means for proper kind of scenes NVlink for 2080Ti is possibly already working！
This is great News!
I will find some other scenes to try.

Jim

[jimho]

I did a quick scene, just to drag in a few people's model (with majorly 8k size map and 40k triangles each) since the 8k map eat Vram a lot, it reaches the Vram limit very soon,
probably the big maps is also kind of one whole piece of thing which is not seperated between GPUs the p2p fails quickly,
then I take out two of the people, adding some trees, they are not as big number triangulars as the carpet and map size is much smaller than the peoples'
Finally IT WORKS！
It's true: the 2080ti is start working with NVLink!

just a quickest scene, all trees and people are overlapped

sure as a start there is limitations and not very practical for a real project, at least this is a great start, the nvlink is a imortant function that we are longing for.

[oscartung]

I've only been able to get very minor speed increase for my scene (around 1% increase) with RTX on.
Still very buggy so far, sometimes scene does not start automatically with rtx on, it just stays on 0/1000, have to click on stop and start.
It also renders incorrectly under some occasion, please see screen shot