RT: Will GPU renderers be able to stack memory with Pascal chip?

[Moriah]

Have been reading a bit about DX12 and how it has support for memory stacking and ability to SLI Nvidia and Amd at the same time. I'm more interested in memory stacking, because if that's possible, GPU renderers will gain a bit boost in terms of performance and current memory limitations. But since i'm no expert, is DX12 even "supported" in Vray RT GPU or even other GPU renderers? Or how does that work?

Will there be any new features in the upcoming Pascal chip that will make GPU renderers a lot faster? Other than the limit of 32GB VRAM.

[Macker]

GPU memory limitations are about the only thing keeping me from recommending that we change from CPU to GPU rendering to my company.

[Moriah]

CPU memory limitations are about the only thing keeping me from recommending that we change from CPU to GPU rendering to my company.

I suppose you mean GPU :P But yeh, same here, even if we can't have all the features that Vray Production has, 80% of our jobs can be done in GPU, it's so much faster and easier to scale.

[Macker]

Yeah I did. Oops, corrected.

[Syclone1]

GPU memory limitations are about the only thing keeping me from recommending that we change from CPU to GPU rendering to my company.

There's that and the fact that not all V-Ray materials and features are supported yet in RT, and Phoenix I'd be on cloud nine for that one lol

[Macker]

As far as I'm aware, everything I use is supported by RT.

[Moriah]

Imo i think they'll be supported sooner or later, also volumegrid, full hair implementation, etc. maybe we should see what Chaosgroup will say in GTC 16 about GPU and stuff.

[vlado]

full hair implementation1

Ornatrix and HairFarm should work fine, I'm not sure we want to deal with Max's Hair&Fur - this is mostly an export thing in Max, not a GPU limitation.

Best regards,
Vlado

[mrhalaby]

hey vlado, I guess a lot of people are wondering if there is any chance that vray will work out-of-core a la redshift3D? or is that discarded/too far in the development line?

[Moriah]

Ornatrix and HairFarm should work fine, I'm not sure we want to deal with Max's Hair&Fur - this is mostly an export thing in Max, not a GPU limitation.

Best regards,
Vlado

Oh, because on the website it says "These plug-ins are only partially supported by V-Ray RT GPU". Only tested ornatrix so far at it seemed ok but nothing special. Could you shed some light on the initial question?

[vlado]

About the initial question, I'm not sure myself yet if this is something that will happen automatically behind the scenes, or something that we need to add support for by manually moving data from one GPU to the others.

Best regards,
Vlado

[savage309]

Have been reading a bit about DX12 and how it has support for memory stacking and ability to SLI Nvidia and Amd at the same time. I'm more interested in memory stacking, because if that's possible, GPU renderers will gain a bit boost in terms of performance and current memory limitations.

Stacked memory is the way Nvidia calls the High Bandwidth Memory (HBM). HBM is new type of memory invented by Hynix and AMD. It uses less power and gives much higher bandwidth (this should allow higher performance). There will be not manual "stacking".

But since i'm no expert, is DX12 even "supported" in Vray RT GPU or even other GPU renderers? Or how does that work?

We don't use DirectX, we use CUDA and OpenCL instead. DirectX/OpenGL is used for realtime graphics.

Will there be any new features in the upcoming Pascal chip that will make GPU renderers a lot faster? Other than the limit of 32GB VRAM.

[/quote]
We don't know enough to say that yet .. but from what I hear - hopefully yes (both for Pascal and Polaris). I am really excited.

You can check this for list of supported features in RT GPU here.

[Moriah]

Thank you very much for the explanation savage let's wait and see then! *crosses fingers*

edit - reading on wikipedia's geforce page it says:
"Unified memory – memory architecture unified so CPU and GPU can access both main system memory and memory on the graphics card"

Unless this is limited to DX12, we could potentially use mboard memory for GPU calculations? Or am i reading it wrong?

[savage309]

GPU memory performs much better and this affects render times. We had some implementations in the past that are targeting this, on which improvements we are working now. Will let you know when we have more on the topic .

[Donfarese]

Can't wait for GPU to fully replace CPU, this is exciting!