RTX and CUDA mixed

[Morne]

Hi guys

If I switch the setting to RTX, for a RTX series card, but I also have a 1080 Ti in there, will the 1080ti also use RTX, or it will fall back to CUDA?

So will both render with RTX or only the one and the other with CUDA?

Will the 1080ti slow things down in this case if I add it to the mix (when setting to RTX mode)?

[Joelaff]

As far as I know RTX does not yet support CPU mode. So to use RTX the machine must have an RTX compatible card installed. So to use the 1080 I believe you would have to use CUDA.

Note the differences between CUDA and RTX are very scene dependent. While RTX is almost always faster it is not always enough faster to justify the inability to use the CPU or machines without RTX cards. For instance, or high core count Threadrippers we see that CUDA mode (using CPU and GPU) is faster than using RTX. So we don't use RTX (especially since few render nodes have RTX cards anyway).

Once RTX works on the CPU as well I think options may open up more.

[a7az0th]

You can use mixed setup to render with RTX.
The RTX codepath is supported on GTX cards, but since they don't have RT Cores, these cards will fall back to software tree intersection.

[Joelaff]

You can use mixed setup to render with RTX.
The RTX codepath is supported on GTX cards, but since they don't have RT Cores, these cards will fall back to software tree intersection.

Oh really! That is cool. So what exactly do you mean by this? Say I have machine A which has an RTX card, and machine B which has a non RTX card. If I set VRayGPU to use RTX and render on both machines what will be used for the rendering? The RTX card on machine A and the CPU on machine B?

Or what?

When I choose RTX on a machine with an RTX card the only device I see is the RTX card. I do not see the option to use the CPU.

In a perfect world it would use the GPU and CPU of every machine, no matter what card type. (Obviously the GTX cards would be slower.)

Also, will the CUDA vs. RTX modes render the same? (identical image?)

Thanks.

[a7az0th]

Joelaff I believe you misunderstood the initial question, or maybe I did
I was saying that if you have 2 GPUs of different generation (GTX and Turing) in a single machine, you can still use the RTX mode of V-Ray GPU and it will render with RTX on both cards. The RTX mechanic provided by NVidia would automatically fallback to an ideantical CUDA implementation that NVidia provides in case your hardware has no RT Cores. So you can use mixed setup (in terms of GPUs) with the RTX mode.

The CPU emulation that we provide however is something that is currently not available in RTX mode and there's a very good reason for that. For CUDA adding this mode was essentially free from a user standpoint, as we already had all the data at our disposal. For RTX this is not the case, as all the acceleration structures are handled, built and used by the RTX engine. This means that in order to emulate this, we will have to duplicate the entire scene in order to maintain 2 separate sets of acceleration structures - one for RTX and one for the emulation device.
This will have a very severe impact on RAM usage and time to first pixel as all the initial work has to be done twice.

As for the last question - CUDA and RTX share identical code so they are expected to render the same.

[Joelaff]

Thanks for the detailed response, Alexander.

Good to know that you can mix and match RTX and CUDA as well in terms of the image output.