OTOY Forums

I think I heard several rumours floating around, but can not find any confirmation.
Can Octane3 use regular built in RAM when it runs out of VRAM?

Hello Stubs

I also had to search around a bit to find the answer to your question.

In the OctaneRender 3 features thread the plugin developer linked you can find one paragraph that seems to be a partial answer to your question:

Moved film buffers to the host and tiled rendering

Source: viewtopic.php?f=33&t=51679

It seems in the early OR 3 alpha 1 &2 builds only the rendered images are moved to the sytem RAM.

- - -

In the OctaneRender 3 announcement thread you can find the following entry:

OpenCL support: OctaneRender 3 will support the broadest range of processors possible using OpenCL to run on Intel CPUs with support for out-of-core geometry, OpenCL FPGAs and ASICs, and AMD GPUs.

Source: https://home.otoy.com/otoy-unveils-octa ... -renderer/


The way I understand this they first may need to integrate OpenCL and then it will be possible to have geometry in system RAM.

There seems to be one "official" answer about OpenCL integration in the OctaneRender™ Standalone 3.00 alpha 1 thread:

Goldorak wrote:

philliplakis wrote:Looks good so far

No OpenCL though?

OSL first, and that is coming in the beta phase.

Source: viewtopic.php?f=33&t=51681&start=10#p257897

- - -


Subjective Commentary:
You have to be aware though that features announced at the GPU conference and the OctaneRender 3 announcement post do not seem to be binding.

It did happen with other announced features that they stopped developing them, released them in another more limited form or postponed them repeatedly.

At this point in time we may just have to wait and see if, how and when they can make geometry in out of core memory work.

- - -

linvanchene wrote:Hello Stubs

I also had to search around a bit to find the answer to your question.

In the OctaneRender 3 features thread the plugin developer linked you can find one paragraph that seems to be a partial answer to your question:

Moved film buffers to the host and tiled rendering

Source: viewtopic.php?f=33&t=51679

It seems in the early OR 3 alpha 1 &2 builds only the rendered images are moved to the sytem RAM.

- - -

In the OctaneRender 3 announcement thread you can find the following entry:

OpenCL support: OctaneRender 3 will support the broadest range of processors possible using OpenCL to run on Intel CPUs with support for out-of-core geometry, OpenCL FPGAs and ASICs, and AMD GPUs.

Source: https://home.otoy.com/otoy-unveils-octa ... -renderer/


The way I understand this they first may need to integrate OpenCL and then it will be possible to have geometry in system RAM.

There seems to be one "official" answer about OpenCL integration in the OctaneRender™ Standalone 3.00 alpha 1 thread:

Goldorak wrote:

philliplakis wrote:Looks good so far

No OpenCL though?

OSL first, and that is coming in the beta phase.

Source: viewtopic.php?f=33&t=51681&start=10#p257897

- - -


Subjective Commentary:
You have to be aware though that features announced at the GPU conference and the OctaneRender 3 announcement post do not seem to be binding.

It did happen with other announced features that they stopped developing them, released them in another more limited form or postponed them repeatedly.

At this point in time we may just have to wait and see if, how and when they can make geometry in out of core memory work.

- - -

I can clear this up. The OpenCL implementation, assuming we don't hit any major roadblocks, will hopefully provide us with a CPU fallback for Octane 3, in which case geometry can be in system memory. Do not expect it to be fast however. If were to do out of core geometry on the GPU, it could slow the speed of GPU rendering to the level of the CPU version, and with a CPU version fallback, we're more or less at this point anyway.

Maybe we could come up with some clever workaround with OOC geometry on the GPU, but I don't think it is worth taking resources from other features to make this attempt right now. Textures in system RAM work well in V2+ with very little speed hit. With the film buffers also OOC on the CPU in V3 (saves GB when using render passes) , and with 2016 cards supporting 32+ GB of HBM VRAM in 2016 (and possibly even more), I think productions targeting 10s of GB of mesh data will be able to do so at full speed on the GPU.

thanks for clearing that up! it never made any sense to me (but what do i know)
so for the moment we'll stick to titan Xs, hoping there will be some awesome announcements from nvidia next year.

I read above... 32GB(!) HBM.. was like, huh?... googled and read some more... "Pascal 10X Faster Than Maxwell"....."3D stacked High Bandwidth Memory"....... sounds like the work of the devil.. I love it.