Can I add the memory from 2 graphics cards together?

[bakayarooo]

I wasn't sure what category to post this under (graphics cards, programming, or operating systems). Sorry if this is the wrong category.

 

I have 2 Titan RTX cards with an AMD Threadripper 3960x and 128GB of ram. I am my OS is Red Hat Enterprise Linux 8 (RHEL8). I have installed the CUDA Toolkit 10.2, NVIDIA CUDA Deep Neural Network library (cuDNN), and NVIDIA Collective Communications Library (NCCL). I am running a 3D-UNet deep learning model. I can run a 224x224x224 volume through my training model with one Titan RTX. I am trying to get run the biggest volume that is possible to get the most detail. I thought that with a 2nd Titan RTX I would double the dedicated graphics memory and be able to handle a larger volume. I have an NVLink bridge. Is there a way to combine the memory from 2 or more graphics cards?

[ericcui]

HI,guy,I'm a newbie too.

[Bombastinator]

Deleted.

Thinking about this I don’t know enough

Edited February 2, 2020 by Bombastinator
Derp

[brain__washed]

Hi there!

 

This seemed very interesting so I went Googling. I know for a fact that under Dx11 and with SLI, you could not add up the memory - the memory between the GPU's would be synced and loaded with the same data afaik.

 

Someone on Reddit had this exact same question, but then regarding videogames. An NVidia rep responded with the following:

 

Quote

VRAM "stacking" is only supported in DX 12's EMA (explicit multi adapter) mode.

When EMA is used, each GPU's memory is accessed individually, carrying only what it needs for its specific task. So, the RAM doesn't "stack" per se, but rather, is assigned different tasks, but that's semantics. In an EMA situation, VRAM on all utilized cards is available.

So to answer your question, yes, VRAM from both of your GPU's can be utilized independently when using DX12's EMA mode. Currently, only one game exists that uses this feature that I'm aware of (Ashes of the Singularity), and I don't think they've fleshed out this mode just yet.

TLDR: No, it doesn't stack regulary.

Source:

 

I'm no expert, but I hope this helps regardless!

[Bombastinator]

40 minutes ago, brain__washed said:

Hi there!

 

This seemed very interesting so I went Googling. I know for a fact that under Dx11 and with SLI, you could not add up the memory - the memory between the GPU's would be synced and loaded with the same data afaik.

 

Someone on Reddit had this exact same question, but then regarding videogames. An NVidia rep responded with the following:

 

Source:

 

I'm no expert, but I hope this helps regardless!

LOL I am vindicated in avoiding comment.  My attitude was “I really don’t think so” but apparently there are instances where it can be done though they are only newly available, and require special programming.  So older apps no, new apps only maybe sometimes.  
 

For what little it’s with good on you for doing the legwork.  I appreciate it