VRAM Pooling will work in all or only supported applications?

[phongle123]

Hello, when pooling VRAM will it work in in general on all applications or will it only be supported in applications that are coded to support VRAM pooling?

 

What happens if it does not support it and you have the 2 GPUs connected. Will it simply only use 1 of the 2 GPUs VRAM or will it error out?

[BobVonBob]

NVLink does VRAM pooling when Nvidia says it can do VRAM pooling. Currently you need the pro workstation/compute cards with two NVLink connectors for VRAM pooling, and probably for Nvidia to flip the magic software switch that lets it happen.

[phongle123]

11 minutes ago, BobVonBob said:

NVLink does VRAM pooling when Nvidia says it can do VRAM pooling. Currently you need the pro workstation/compute cards with two NVLink connectors for VRAM pooling, and probably for Nvidia to flip the magic software switch that lets it happen.

Didnt know that you couldnt pool VRAM yet. It will also work for the TITAN RTX. It says so on the official product page. 

 

Since VRAM Pooling is not a real thing yet then I guess no one besides nvidia knows the answer to this question. 

[Mira Yurizaki]

I would imagine it should be transparent to the application programmer and what they see is the combined total. However, they would still have to be aware of certain use cases that may cause performance issues since while NVLink is faster than PCIe 3.0 x16, it's still much slower than VRAM speeds. So if anything, all this would allow you to do if you were looking at this from a games perspective allow for a larger texture/asset pool since I imagine passing around render targets would be a bad idea over a slower bus.

[KarathKasun]

VRAM pooling is only available on Quadro level cards.  The application also has to be programed to support it.

 

Due to these limits, it will only be supported in CAD, offline rendering, and compute apps more than likely.

[phongle123]

36 minutes ago, KarathKasun said:

VRAM pooling is only available on Quadro level cards.  The application also has to be programed to support it.

 

Due to these limits, it will only be supported in CAD, offline rendering, and compute apps more than likely.

I already stated it works in Titan RTX since it states it on the official site. The whole point of why I created this thread is for offline rendering. Additionally, I wouldn't want to spend 5k on GPUs to not have an application support VRAM Pooling and according to @BobVonBob VRAM Pooling isnt even available right now. So, idk if VRAM pooling works out of the box for everything as in I'll have 2x the VRAM all around or it will only work in specifically coded softwares.

 

50 minutes ago, Mira Yurizaki said:

I would imagine it should be transparent to the application programmer and what they see is the combined total. 

From what I'm getting you're saying that the program/software should simply receive a minor update to adjust to VRAM Pooling to shows the 2x VRAM of a pooled NVLink SLI Card? This won't be for gaming so I don't much care for those results. 

[KarathKasun]

4 minutes ago, phongle123 said:

I already stated it works in Titan RTX since it states it on the official site. The whole point of why I created this thread is for offline rendering. Additionally, I wouldn't want to spend 5k on GPUs to not have an application support VRAM Pooling and according to @BobVonBob VRAM Pooling isnt even available right now.

 

From what I'm getting you're saying that the program/software should simply receive a minor update to adjust to VRAM Pooling to shows the 2x VRAM. This won't be for gaming so I don't much care for that results. 

https://www.chaosgroup.com/blog/what-does-the-new-nvidia-rtx-hardware-mean-for-ray-tracing-gpu-rendering-v-ray
 

V-Ray supports it.

[Mira Yurizaki]

6 minutes ago, phongle123 said:

From what I'm getting you're saying that the program/software should simply receive a minor update to adjust to VRAM Pooling to shows the 2x VRAM. This won't be for gaming so I don't much care for that results. 

If I were a hardware manufacturer, that's how I would make it look like to the application. It would seem silly for the application developer to have to manage and worry about a low level detail like this. And if I ask them to be aware of it and code specifically for it, unless they're a huge customer and they need the performance, they're probably not going to care. The only thing that the application developer would need to be aware of are those certain situations where the slower NVLink could impact performance, but that's a matter of profiling the application and avoiding scenarios that lead to the hardware doing that. Basically, it'd be no different than multi-socket CPU systems. Applications see the system with a single pool of RAM (even though there's two or more) and a bank of logical processors.

[KarathKasun]

V-Ray also supports pooling with 2080/2080 Ti.  However, the NVLink bandwidth is low enough that using it slows things down a little if you do not need the extra VRAM capacity.

 

https://www.chaosgroup.com/blog/profiling-the-nvidia-rtx-cards