GPU setup for RT

[Christian_Garimberti]

Hi guys,

watching the 2 Deep dive extended video on the YouTube channel

https://www.youtube.com/user/VREDPRO

I see that Simon and Danny have always a GPU for denoising and 2 or more gpus for rendering.

Supposing I have a workstation with 3 gpus:

1. Do I need more then 1 pro license also if use only 2 gpu for rendering?
2. should I use homogeneous gpus? Or can I put the one I use for denoising different from the other 2?
3. If I can’t have a workstation wit 3 gpus but I have two workstation, the master with a gpu for denoising ant the other with the 2 gpu as a cluster for rendering, should I have more or less the same performances of a single one?

Best

Chris

[michael_nikelsky · Accepted solution]

They use two systems and cluster rendering, a master node, which does the denoising, and a RTX Server, which does the rendering.
With local rendering the denoising is always run on the main GPU that is also responsible for the image output and you don´t have control about that. You can only limit the GPUs to use in the preferences but if you don´t put the GPU there where the OpenGL context is you will just see a black screen.

For 3 GPUs you would need an additional VRED Core license for the third GPU (or limit the usage with the preferences setting I mentioned).

Performance is a difficult topic since it also depends a lot on you LAN connection. In general Cluster rendering will be slower if the same number of GPUs are used simply because there is a lot of additional work involved for network communication, load balancing between nodes and so on. Also a GBit LAN is really the lowest possible connection you should have, 10GBit or infiniband are highly recommended for clustering due to the amount of data that needs to be send. The benefit of using a cluster is to use more GPUs than can fit into a single system. If you don´t have that you should stick to a single system. 

Regarding homogenous GPUs: In a single system the GPUs should be homogenous, we distribute the work so every GPU gets about the same amount of work so in a heterogenous system the slowest GPU will slow down all the others.

In a cluster environment there is a limited amount of work balancing between cluster nodes but it is still recommended to use the same GPUs to avoid issue due to different memory sizes or architecture. The only exception is the master node which does the denoising, this could also be a different GPU.

Kind regards

Michael

[Christian_Garimberti]

Hi Michael, thank you for all the clarifications.

Best

Chris