Multiple GPU render speed

[abdul_rahman]

Hi,

I'm going to upgrade with another GPU. Currently i have RTX 3070, and having another RTX 3090 will increase the render speed?

[npg]

Yes, rendering on two GPUs is generally faster. But note that there is a maximum achievable framerate when using two GPUs. This is because the second GPU has to transfer data to the main GPU on each frame. This happens over the PCI-express bus which is not very fast (PCI-e v4 is better than PCI-e v3 obviously). But this only matters if your scene is not too complex and a single GPU can render it at a very high framerate. Otherwise, if the framerate on a single GPU was already low due to the complexity of the scene, adding a second GPU should make it close to twice faster if it's the same model. In other words if it takes more time to render one frame than the time it takes to transfer the rendered pixels back to the main GPU, you are not affected by the PCI bottleneck.

The other caveats that apply are that:
- Light Cache is not supported on multiple GPUs. This should not be a big deal, it should still be faster to render on two GPUs without LC than on one with LC.
- Each GPU holds all the necessary scene data, so you're limited by the amount of memory on the smaller GPU. The main GPU uses a little bit more memory for display buffers, so it's better if you use the bigger card as the main/display card.

[q976431]

Yes, rendering on two GPUs is generally faster. But note that there is a maximum achievable framerate when using two GPUs. This is because the second GPU has to transfer data to the main GPU on each frame. This happens over the PCI-express bus which is not very fast (PCI-e v4 is better than PCI-e v3 obviously). But this only matters if your scene is not too complex and a single GPU can render it at a very high framerate. Otherwise, if the framerate on a single GPU was already low due to the complexity of the scene, adding a second GPU should make it close to twice faster if it's the same model. In other words if it takes more time to render one frame than the time it takes to transfer the rendered pixels back to the main GPU, you are not affected by the PCI bottleneck.

The other caveats that apply are that:
- Light Cache is not supported on multiple GPUs. This should not be a big deal, it should still be faster to render on two GPUs without LC than on one with LC.
- Each GPU holds all the necessary scene data, so you're limited by the amount of memory on the smaller GPU. The main GPU uses a little bit more memory for display buffers, so it's better if you use the bigger card as the main/display card.

Directly buy two graphics cards and plug them into the motherboard to complete the installation of dual graphics cards?
Or do I need to pay attention to what can make vatange support it

[npg]

I'm sorry, I don't understand the question very well.

Vantage supports any combination of compatible GPUs (NVidia RTX). It's best if the two are the same or similar in speed and RAM. It's strongly recommended to have both on full-speed 16x lane PCI-e slots.

[q976431]

I'm sorry, I don't understand the question very well.

Vantage supports any combination of compatible GPUs (NVidia RTX). It's best if the two are the same or similar in speed and RAM. It's strongly recommended to have both on full-speed 16x lane PCI-e slots.

How to set up the primary video card for vantage
I have two video cards, 3080 and 2080 Ti
I tried to set or switch the video card socket, and he always set the 2080 Ti as the primary video card, but obviously the 3080 will be faster, I don't know on what basis? Is it because the 2080 Ti 12gb, 3080 10gb are automatically set depending on the size of the video memory?
Please provide me with a parameter that can set the main video card, because I need to use 3080 as my main video card, but I have found all the ways now, but I can't.

[q976431]

How to set up the primary video card for vantage
I have two video cards, 3080 and 2080 Ti
I tried to set or switch the video card socket, and he always set the 2080 Ti as the primary video card, but obviously the 3080 will be faster, I don't know on what basis? Is it because the 2080 Ti 12gb, 3080 10gb are automatically set depending on the size of the video memory?
Please provide me with a parameter that can set the main video card, because I need to use 3080 as my main video card, but I have found all the ways now, but I can't.

I'm confirming my guess. Does vantage judge the primary and secondary graphics cards by identifying the size of the graphics memory
When I changed 2080ti (12g) to 2080s (8G), my 3080 (10g) became the main graphics card
But there is still a problem.
When I only use 3080, the default scenario is 100fps
But when I use 3080 + 2080s. The default scenario is only 45fps Even if the 3080 has been used as the main graphics card

[npg]

This is correct, we sort the devices by memory size. Before comparing framerates, disable "Noise threshold" by setting it to 0 in the render panel. Otherwise the framerate will change a lot.

As I said, running dual GPUs limits the maximum framerate. On simpler scenes it may be lower because a single card does not have to communicate with another GPU and therefore achieves higher FPS.

[q976431]

This is correct, we sort the devices by memory size. Before comparing framerates, disable "Noise threshold" by setting it to 0 in the render panel. Otherwise the framerate will change a lot.

As I said, running dual GPUs limits the maximum framerate. On simpler scenes it may be lower because a single card does not have to communicate with another GPU and therefore achieves higher FPS.

Thank you very much for your answer.
Will you consider allowing customization of which one is the main graphics card? Because more memory doesn't mean better performance.
In addition, I would like to ask about whether multiple graphics cards are shared display memory? For example, 3080 (10g) + 2080 (8g) = 18 GB of available memory, or it is limited to the lowest display memory (8g)?
This is very important, because this means I use 3090 (24G) with 3080 (10g), it is ultimately the maximum number of 34G memory, or is it limited by 3080 to the maximum no more than 10g?Because I seem to see that you mentioned that each graphics card will load scene data, it means that 20G display memory is loaded on the main graphics card 3090, load failed on the auxiliary graphics 3080, and finally rendering failed?

[Alexander.Atanasov]

Hi q976431 ,

You can set which GPU to be main from Edit > Render devices dialog:


Uncheck "Select automatically" and reorder the GPUs with the arrow keys on the right.

Best regards,
Alexander

[q976431]

Hi q976431 ,

You can set which GPU to be main from Edit > Render devices dialog:


Uncheck "Select automatically" and reorder the GPUs with the arrow keys on the right.

Best regards,
Alexander

i see!!!!!
Thank you.

This UI seems to have a bug
When I moved the 2080 graphics card to the first place, made it the main graphics card and restarted vantage
The main graphics card became 2080, but when I opened this UI option again, I found that he did not rank the 2080 graphics card first, but returned to the original state, although the 2080 has become the main graphics card.
This means that when I want to switch back to 3080, I have to move the first 3080 graphics card to the end, return it to the first place again, and finally click apply

Normally, when I set the main graphics card to 2080 and restart, the UI should display 02
But now it's 01
This will not affect my setting up the main graphics card
But it's really a bug about UI code.
The latest UI sort should be refreshed for the user

But I still have questions about the memory of the two GPUs

[Alexander.Atanasov]

Chaos Vantage currently does not share GPU memory meaning that 3080 (10g) + 2080 (8g) will not equal 18 GB. The maximum GPU memory that can be used is equal to the smallest GPU, in this case 8gb. We have already have this request and we will look into it but cannot give ETA at the moment. We will keep you informed when there is something in development.

Best regards,
Alexander

[q976431]

Chaos Vantage currently does not share GPU memory meaning that 3080 (10g) + 2080 (8g) will not equal 18 GB. The maximum GPU memory that can be used is equal to the smallest GPU, in this case 8gb. We have already have this request and we will look into it but cannot give ETA at the moment. We will keep you informed when there is something in development.

Best regards,
Alexander

Thank you very much. This let me know that the 3090 + 3080 scheme may not be feasible. It should be limited by the maximum display memory from the 3080 graphics card
For example, your rtx8000 will be limited by the 8g of rtx2080?

[chromaloma]

Chaos Vantage currently does not share GPU memory meaning that 3080 (10g) + 2080 (8g) will not equal 18 GB. The maximum GPU memory that can be used is equal to the smallest GPU, in this case 8gb. We have already have this request and we will look into it but cannot give ETA at the moment. We will keep you informed when there is something in development.

Best regards,
Alexander

Hi, where are you in this development?
I am impressed by how Vantage is turning our workflow into something better, and it might "render" our CPU renderfarm as void soon.
Right now we have 1x3090 in our configuration, and that works on MOST of our scenes, but I am eager to learn more about what exactly improves in a dual GPU setup.
My consideration would be to have the firm invest in 2x4090, or 1x 8000, just for the sake of VRAM.
Will VRAM and CUDA cores ~double up in a 2x GPU setup in near-future release?
Would love to see some benchs on this.

[npg]

No, this is not in development. NVIDIA deprecated NVLINK, so the hardware we need to implement memory sharing now only exists on old GPUs, so there's no point pursuing this.

If you use 2 GPUs each one will have its own copy of the scene data and they will split the render work (except for denoising - this is done on the main GPU). So speed can be up to 2x if you're not denoising (or "final pass only"), but realistically somewhere below 2x. Using two GPUs also puts a cap on the maximum framerate, depending on resolution - we can't render faster than the time it takes for the second GPU to send its result back to the main GPU. So PCI-e 4+ is recommended.