Nvidia 3000 series support Tweet

[Goldorak]

You will need 2020.1.5 out this week to use ampere at near full speed. Further ampere optimizations will continue in 2020.2 as we get more time to tune to the arch.

[Goldorak]

RTX on is way faster, yes

That's not what we are asking.

Reframing question:
Is RTX Gen 2 a unique and new technology that will require time for Otoy to re-code to utilize...or is it simply a performance improvement where it is a seamless/quick transition?

Same question for Tensor core features (denoising/upsampling/AI lights). Is 3xxx just faster, or bring something that is technologically different, requiring Octane re-writes that take time and delay user benefit?

Nothing special other than perf. tuning on RT/TC cores as we have done on cuda cores before

[Iceman9]

RTX on is way faster, yes

That's not what we are asking.

Reframing question:
Is RTX Gen 2 a unique and new technology that will require time for Otoy to re-code to utilize...or is it simply a performance improvement where it is a seamless/quick transition?

Same question for Tensor core features (denoising/upsampling/AI lights). Is 3xxx just faster, or bring something that is technologically different, requiring Octane re-writes that take time and delay user benefit?

Nothing special other than perf. tuning on RT/TC cores as we have done on cuda cores before

Thank you for your reply. Cheers.

[Timmaigh!]

The RTX speed up is scene dependent, so while 2019 havana scene, octane 1,2 and 3 splash screen scenes get 2-4x speed boost, that is not the case with the the simpler octane bench scenes like idea which is a light bulb in a room, and all shading, little tracing we can push to RT.

For now, turning on RTX, getting a 20% OB boost with the existing OB4 scenes actually makes sense because it is a good measure of what almost any scene is likely to do on RTX - after all the OB4 scenes render identically with RTX on, and served as our benchmark for 5 years before RTX was a factor.

That said, the RTX on boost on Ampere for these scenes may not be the same as Turing, making classic OB potentially higher with RTX on on Ampere. We’ll have to wait until that is all public.

Beyond standard OB, we have done deep testing 40+ heavy real world scenes both on ampere, turing et al. with RTX on and off, so we now have a very good sense of the real world averages relative to classic OB. We are planning to have a future OctaneBench add more if these heavier scenes, which will likely push OB higher on Ampere/Turing relative to the 100 OB set on a 980.

We’ll share these preliminary results as well once the cards are public. But, rest assured, the official OB scores you will see for Ampere are really impressive relative to last gen, either with RTX on or off,
classic OB scenes or with new OB candidate scenes.

Many thanks for your answer. Looking forward to more info.

[voon]

Does the 3090s NVLink mean, Octane can pool VRAM across several cards? I'm confused there. So two 3090s would have 48 GB of scene memory, while two 3080s only have the old, non shared 10 GB?

[coilbook]

Hi Otoy,
Will you take advantage of rtx io where you bypass CPU and load data from ssd directly to vram https://youtu.be/QKx-eMAVK70?t=1434 ?

[Timmaigh!]

And one more question, since Octane X now supports AMD hardware, should we expect regular Windows Octane with AMD support as well?
We are all looking at upcoming Ampere cards, but AMD is about release their new stuff as well, with their own hardware raytracing units, rumored massive cache and so on... for all we know it might turn out actually better performing than Nvidia...and cheaper too. If Octane happens to support it now, would be nice to know in advance.

[dobe]

I wouldn't hope that the graphics cards from AMD will be better than those from Nvidia. - Especially because of count of cuda cores.

the question for me, can i go with two 3080 and break the VRAM limit (10GB) by using the out of core option in octane?

[Timmaigh!]

I wouldn't hope that the graphics cards from AMD will be better than those from Nvidia. - Especially because of count of cuda cores.

the question for me, can i go with two 3080 and break the VRAM limit (10GB) by using the out of core option in octane?

I dont think the cuda core number is the only important spec. Somehow, AMD upcoming GPU architecture may have some othe properties, which may bode surprisingly well for raytracing/pathtracing/Octane, like those bigger caches and whatnot. Or their single "core" may be more powerful and performant than Nvidia´s one.
I am not necessarily expect it to happen, but if hypothetically Octane happens to support it and the AMD card performs about the level of 3080, while having 16GB of RAM and be 100 cheaper, would you not at least consider it?

That 3080 question was not directed at devs, but i guess i can answer it as well. You will have 10GB at your disposal with 2x 3080 (or better said 80 percent of it cause of CUDA app related VRAM limit), after that Octane will surely use Out of Core (system RAM). The 3080 does not support NVlink, so no "memory pooling" feature. If that is what you asking.

BTW, regarding Out of Core, i have question of my own. Say you combine 24GB 3090 and 11GB 2080Ti (or 10GB 3080 or whichever GPU with smaller VRAM). Say you have a scene 18GB large. Does Octane now use 11GB of VRAM on both cards (including 3090, despite bigger VRAM) and then fills the rest with OOC, or does it use 18GB on 3090 and OOC kicks in only for the 2080Ti?

[dobe]

That means the "

Code: Select all

out of Core

" function in Octane does only work in combi with NVlink? right?