general GPU resource question Tweet

[protovu]

Hi Simon,
I have a general information question regarding GPU Memory.

From the attached, you can see that my Used GPU Memory does come close to my Available GPU Memory.
Is there a setting that I need to look at i order to access more GPU Memory?

Also, shouldn't "Used" + "Available" = "Total" ?

Thanks for your help,
Rick

[Sighman]

The amount of memory you have available is based on the GPU with the least amount of memory. So if you have say 3 GPUs, 2 with 4gb and 1 with 2gb and you enable rendering on all 3 GPUs then you would only see 2gb available. If you disabled the 2gb gpu then you would see 4gb available.

Also, the operating system will take up a chunk of memory on the gpu attached to the monitor. This memory will not be available for rendering. This is why you see total at 4gb but only 3.5 available for rendering. If you disable the gpu attached to the monitor the amount available will go up. Octane also needs a smallish chunk for its code so available plus used will always be less than total.

I hope that makes sense.

[protovu]

The amount of memory you have available is based on the GPU with the least amount of memory. So if you have say 3 GPUs, 2 with 4gb and 1 with 2gb and you enable rendering on all 3 GPUs then you would only see 2gb available. If you disabled the 2gb gpu then you would see 4gb available.

Also, the operating system will take up a chunk of memory on the gpu attached to the monitor. This memory will not be available for rendering. This is why you see total at 4gb but only 3.5 available for rendering. If you disable the gpu attached to the monitor the amount available will go up. Octane also needs a smallish chunk for its code so available plus used will always be less than total.

I hope that makes sense.

I think I understand in terms of what is shown as "Available". Not sure I understand in terms of "Used" for the actual render, or if I can increase the "Used" in my render set up to actually speed things up.

If I disable the smaller GPU , as you suggest, would I be doing so via the attached graphic (Nvidia Dialog Box) , or is there some other recommendation, like plugging both monitors into the motherboard (not even sure if this is possible)? Or, are you saying that if I uncheck the "Active" selection for the GTX970 via the Octane Render for Carrara Settings/Devices dialog box, my performance (Used + Available) will go up even though I am utilizing fewer GPUs?

Just curious, does "Used" refer to the limit of what is actually being by Octane during the Octane render, or what is used by the system overall?

Thanks for your help,
R

[Sighman]

Performance is obviously based on the number of gpu cores you have set to 'active' on the devices tab. Available memory, on the other hand, is based on the smallest 'active' gpu. More memory will not give you better performance. it will, however, allow you to load bigger scenes. Used is the amount of memory that octane has obtained in order to render your scene. It contains your geometry, textures, and materials plus any buffers that octane needs to do its thing. Used does not include any memory needed by the OS or the Octane kernel. As long as your scene fits in available memory you can render your scene. Any extra memory is idle. If you are tight on memory then you can enable out-of-core textures. This will allow octane to load textures in main memory but it does have a significant performance impact. It is always better if you can fit your entire scene in the gpu. From the example screenshot you posted your scene easily fits onto all three of your GPUs so there is nothing more you can do from a gpu resource point of view to maximize performance.

[protovu]

Performance is obviously based on the number of gpu cores you have set to 'active' on the devices tab. Available memory, on the other hand, is based on the smallest 'active' gpu. More memory will not give you better performance. it will, however, allow you to load bigger scenes. Used is the amount of memory that octane has obtained in order to render your scene. It contains your geometry, textures, and materials plus any buffers that octane needs to do its thing. Used does not include any memory needed by the OS or the Octane kernel. As long as your scene fits in available memory you can render your scene. Any extra memory is idle. If you are tight on memory then you can enable out-of-core textures. This will allow octane to load textures in main memory but it does have a significant performance impact. It is always better if you can fit your entire scene in the gpu. From the example screenshot you posted your scene easily fits onto all three of your GPUs so there is nothing more you can do from a gpu resource point of view to maximize performance.

Thank you, Simon. Sorry for the elementary questions, but:
So we are distinguishing "memory" as "GPU memory" here - as distinct from " GPU cores"? I was conflating the two.

With your points above in mind, is there a way to predict performance enhancement relative to an added video card, say another Titan? Some formula I can look at regarding GPU cores, perhaps?
I hate to spend the additional if it yields only marginal results.

[Sighman]

Adding additional cores to the render process is almost linear when it comes to performance. If you have 6000 CUDA cores and you add an additional 3000 CUDA cores then you will get a ~50% improvement in performance. I would not look at adding a Titan, although this is a great card it is expensive and Octane does not take advantage of its 64bit floating point architecture. You would see much better returns adding a GTX980ti. You could add a GTX1080 but Octane is not optimized (yet) for the Pascal architecture. I would stick to Maxwell cards for now (unless you want to get into VR on the side). The GTX980Ti has 2816 CUDA Cores and 6GB of GPU memory for $959.62US. A Titan X had 3072 CUDA cores and 12GB of memory for $1,834.55US. You do the math...

[protovu]

Thank you, Simon. Terrific information.
Best,
Rick